International Journal of Artificial Intelligence in Education (2000), 11, 344-376 System Intelligence in Constructivist Learning Fabio N. Akhras and John A. Self Computer Based Learning Unit, University of Leeds, Leeds, LS2 9JT, England E-mail: akhras@pcs.usp.br, J.A.Self@cbl.leeds.ac.uk Abstract. The aim of this paper is to present a perspective on intelligent systems to support learning that is in line with constructivist views of learning. In order to develop such a perspective we have defined formal mechanisms to support knowledge representation, reasoning, and decision making in intelligent systems, that are attuned to the values of constructivist views of learning. These point to the importance of the context of learning, stress that learning involves active interaction, and emphasise the process rather than the product of learning. The theoretical models that constitute our approach enable intelligent learning environments to evaluate learning according to four properties of constructivist learning processes: cumulativeness, constructiveness, self-regulatedness, and reflectiveness, and to make decisions about the learning opportunities to be provided to the learners, taking into consideration the affordances of learning situations regarding these properties. The approach has been implemented in INCENSE, which is an intelligent learning environment in the domain of software engineering. INTRODUCTION Constructivist theories of learning emphasise an active and autonomous role for the learners to construct their own understanding through interacting in an environment in which the knowledge of the domain is not explicitly separated from the context in which it applies. The focus is on the process through which the learners experience the environment and interpret their experiences rather than on the acquisition of a previously defined target domain knowledge. These emphases of constructivism point to a general shift in focus from teaching to learning and bring to the fore a set of issues that differ in fundamental ways from the issues that have been addressed in the design of intelligent systems to support learning. Intelligent systems to support learning have emphasised the use of artificial intelligence in education with three main purposes: representation of the knowledge to be learned, inference of the learner's state of knowledge, and planning of instructional steps to be followed by the learner. The focus of these systems on the explicit definition of a model of the domain knowledge to be acquired by the learner, and on modelling the learner's knowledge state in terms of the learner's correct knowledge or misconceptions, which are used as a basis to evaluate learning and guide instructional interventions, seemed difficult to reconcile with constructivist views of learning, and have led researchers engaged in the development of computational support for constructivist learning to move away from the idea of using intelligent systems to provide this support (Derry and Lajoie, 1993; De Corte, 1995). The general view, as stressed by Kommers, Lenting and van der Veer (1996), is that constructivism indicates "a trend towards more autonomy for the learner, instead of an ever increasing cybernetic sophistication of so-called 'system intelligence' in tutoring programs" (p. 408). However, it may be that it is not the idea of a "system intelligence" that is antithetical to constructivist forms of learning but the particular kind of system intelligence that has so far been designed in intelligent systems to support learning. Currently at the Department of Computer Engineering, Polytechnic School, University of São Paulo, CP 61548, 05425-970, São Paulo, SP, Brazil (e-mail: akhras@pcs.usp.br). 344 System Intelligence in Constructivist Learning For instance, it is clear that the focus of intelligent systems to support learning in terms of knowledge representation, reasoning, and decision making, as discussed above, reflects the values of the particular view of learning that is emphasised in these systems in regard to the nature of knowledge, the way learners learn, and the way learning can be promoted. Therefore, alternative views of learning, such as constructivism, may similarly benefit from a system intelligence in which the mechanisms of knowledge representation, reasoning, and decision making, originate from a formal interpretation of the values of that view of learning. As a consequence, the resulting intelligent behaviour of the system will (by definition) not be in contradiction with the values emphasised by that view of learning, as it appears to be today where there is a tension between the underlying values emphasised by intelligent systems to support learning and the values emphasised by constructivist views of learning. Therefore, what is needed is the development of a different kind of system intelligence that is based on methods of knowledge representation, reasoning, and decision making, better attuned to the values of constructivist views of learning. For example, given the shift in focus offered by constructivist views, from the product to the process of learning, the issue of evaluating learning which is central to the individual adaptation of the learning experiences to the learner's perceived needs, shifts away from a model of "what" is learned into a model of "how" knowledge is constructed. In this paper we discuss some of the main issues that concern constructivist theories of learning, and provide a theoretical, computational basis for addressing these issues in the design of a system intelligence to support constructivist forms of learning. The paper is organised as follows. After the discussion of constructivist issues in the next section, the following section outlines the implications of these issues to the design of a system intelligence. The following four sections present our main theoretical developments related to the issues of context, interaction, process and affordances. Then, the next section describes INCENSE, an intelligent system to support learning of software engineering concepts. This system, implemented in Prolog, reasons about context, interaction, process and affordances using the formalisms presented in the theoretical sections. For example, as it interacts with the student it builds up a picture of the affordances of potential situations to the student connected with that particular interaction. The final section presents the conclusions. A CONSTRUCTIVIST VIEW OF LEARNING A view that is emerging from constructivist theories of learning emphasises four aspects as holistically coexisting in any learning process: 1. Context - an essential part of what is learned is the situation in which learning takes place, which refers to the physical as well as to the social environment in which the learner is engaged in activity, and might include physical entities, tools, and other people. 2. Activity - all knowledge is constructed by the learners through actively interacting in situations in which they experience a domain and interpret their own experiences. 3. Cognitive structures - previously constructed knowledge influences the way learners interpret new experience and affects their thinking and acting. 4. Time-extension - the construction of knowledge occurs over time from the learners' attempts to connect their previously developed experiences to the new ones. To take these four aspects into consideration in a holistic way means to assume an inseparability between context, physical and psychological phenomena, and the flow of experience, in order to understand learning. It implies a focus on the relationships that develop between these four aspects in a process of learning, rather than on their independent characteristics. Support for this view is found in recent research on education that has pointed to the need for developing theoretical frameworks in which psychological and environmental aspects are 345 Akhras and Self integrated. For example, Vosniadou (1996) suggests that research is needed to improve our understanding of how cognitive processes and structures interact with environmental variables. According to her: "cognitive psychology provided rich descriptions of what is learned but failed to provide fruitful hypotheses about how learning happens and more specifically about the environmental variables that influence the knowledge acquisition process" (p. 104). This tendency towards the study of psychological and environmental aspects of phenomena in an integrated way is also in line with what has been proposed as a transactional perspective for research and theory in psychology (Altman and Rogoff, 1987). According to this perspective, however, the focus is not only on the relations between individuals and their environments, but also on the temporal qualities of these relations considered as inherent aspects of phenomena, and embodying the flow and dynamics of the individual's relations to social and physical settings. The Focus on Interaction The focus on the relations between individuals and their environments aims to stress that, according to constructivism, learning is essentially interactive. Knowledge (or knowing) does not arise solely from the entities of the environment nor from the learner but from the interactions between them. A fundamental consequence of this is that individual cognitions can only be explained in terms of their contributions to interaction (Greeno, 1997). In addition, the entities of the situation in which the learner is interacting, i.e., the meanings of these entities for the learner, can also only be explained in terms of their contributions to interaction. In fact, when learners are interacting with these entities they are not interacting with entities "as they really are", but rather dealing with their previously constructed perceptual and conceptual structures (von Glasersfeld, 1996). This means that the context of a learner's experience is a flexible notion whose meaning is subject to the learner's interpretation. Therefore, to understand the way in which interaction influences learning we need to understand the ways in which environmental properties (or their interpretations) and properties of the individual cognitive structures contribute to interaction. To characterise these contributions Gibson (1977) proposed the notion of affordances, which refer to things in the environment that can contribute to interaction taken with reference to an individual. One of Gibson's examples is the postbox that affords letter-mailing to a letterwriting human in a community with a postal system. In addition, Greeno, Moore and Smith (1993) propose that an ability for a particular kind of interactive activity is what enables an individual to engage in interactions of particular kinds in a situation. According to this view, affordances and abilities to interact are relative to each other, i.e., a situation can afford an interactive activity for an individual who has appropriate abilities, and an individual can have an ability for an interactive activity in a situation that has appropriate affordances. Neither an affordance nor an ability is specifiable without considering the other (Greeno, 1994). This illustration, given by the notions of affordance and ability, of the way in which interaction can be holistically shaped by aspects such as: the context in which interaction takes place, the activity developed by the learner, and the cognitive structures of the learner interacting in the context; indicates the need to better understand the roles that these three aspects of interaction play in learning. It also indicates that the way these three aspects are intertwined might give rise to very complex issues. For example, Saada-Robert and Brun (1996) have pointed to studies that show that even acquired knowledge is not simply applied to a situation but reconstructed according to the structure of the situation. Or, according to Brown, Collins and Duguid (1989), a constructed concept "will continually evolve with each new occasion of use, because new situations, negotiations, and activities inevitably recast it in new, more densely textured form" (p. 33). 346 System Intelligence in Constructivist Learning The Role of Context Concerned with the role of the context in learning interactions, researchers have investigated the differences between the kinds of learning activity that take place in school and the ones that take place in everyday real life and work situations (Resnick, 1987; Brown, Collins and Duguid 1989), and have suggested that in order that students become able to think with and about the entities of a domain, rather than just learning what these entities are, students need more than abstractions and self-contained examples. They need to learn how these entities are generated and how they work in authentic activities (Greeno, 1989; Brown et al. 1989). According to Bednar, Cunningham and Perry (1992), one of the practical consequences of these ideas to the design of learning situations is to focus on portraying tasks that take into consideration what real people typically do in real life contexts where knowledge domains are not readily separated and information from many sources as well as varied perspectives are necessary. The Role of Activity Concerning the role of activity, a basic premise of constructivism is that all knowledge is subordinated to action. According to Piaget and Garcia (1991), there are two aspects that characterise the meanings of objects. First, it is the action of utilising objects, or "what can be done" with the objects either physically, such as moving or breaking them into pieces, or mentally, such as classifying or relating them. Second, it is the action of constructing objects, or "what the objects are made of". As for the meanings of actions themselves, they are characterised by "what the actions lead to" in the transformations they produce on objects or situations. It follows that the meanings of objects are then characterised by the particular activities in which these objects are utilised or constructed, and by the particular situations in which these activities take place. The Role of Cognitive Structures Concerning the role of cognitive structures in learning interactions, a fundamental implication of a view of learning that emphasises the active participation of the learners in constructing their own knowledge from the activities that they develop in situations is that, in this process, everything is subject to the learners' interpretation. Situations and activities do not have an objective reality but rather reflect what the learners are "able to fit" into the cognitive structures that they already have, which correspond to their prior knowledge. A key issue in this process is what is meant by "able to fit", which in its positive sense is related to the issue of transfer - when knowledge learned in one situation is used later in another situation; while in its negative sense has to do with the problem of inert knowledge - failure to use in one situation relevant knowledge learned before. In general, according to Greeno, Moore and Smith (1993) the issue involves an understanding of "how learning to participate in an activity in one situation can influence (positively or negatively) one's ability to participate in another activity in a different situation" (p. 100). The Focus on Process The focus on interaction has implied that we should take into consideration in the analysis of learning phenomena the properties of the interactions that develop from the learners' activities in the physical and social contexts of their environments, rather than isolated aspects of learners' cognitive structures, learners' activities, or contexts in which the interactions take place. It suggests that learning can be better understood from the circumstances provided by the relationships between the learner's activity, the context in which the activity develops, and the cognitive structures that the learner brings to the activity. 347 Akhras and Self In addition, the focus on the temporal qualities of aspects of interaction implies that we should take into consideration the properties of the relations that develop over time between aspects of single interactions in a process of learning. It suggests that an interaction, and consequently, the learning that derives from it, can be better understood from the circumstances provided by the flow of experience that connects that interaction to other interactions located in different times. Therefore, a fundamental issue for research is to understand the meaningful ways in which aspects of an interactive learning experience in one situation can connect to aspects of an interactive learning experience in another situation, in a course of interaction between learner and environment, which will characterise ways of developing process-related properties of constructivist learning, such as the properties of being cumulative, constructive, self-regulated, and reflective, that have been described by Shuell (1992) and Simons (1993), among others. These properties will be considered in detail later. Pedagogical Situations According to the constructivist view of learning presented in the previous sections, learning may result from time-extended processes of interacting in situations. However, not all kinds of interaction in situations lead to the same sort of learning and some interactions may not lead to any learning at all. Indeed, the discussion about the roles of context, activity and cognitive structures, indicates that different situations for different learners, or for the same learner at different times, may lead to different kinds of learning. Similarly, the process that emerges from the way successive interactions in situations are chained over time, may result in different flows of learning experience for different learners. Based on the notion of affordance, conceived by Gibson (1977), we can say that the utility of a situation for a learner at a certain time is determined by the affordances of that situation with respect to features of single interactions (involving relations between context, activity and cognitive structures) and with respect to features of time-extended processes of interaction (involving relations between single interactions). As Resnick (1996) points out, "learning and development occur when individuals prepared for certain concepts encounter environments with the kinds of affordances they need to elaborate these prepared structures" (p. 39). Therefore, a pedagogical situation, i.e., a situation that can provide learning opportunities for a particular learner at a particular time, shall afford interaction in contexts that embed opportunities for activities, for learners capable of recognising and acting in the situation in ways that can develop further their cognitive structures. As for time-extended processes, a pedagogical situation shall afford certain interactive experiences - involving particular aspects of context, activity, and learner's cognitive structures - that allow the development of relations over time with aspects of interactive experiences developed by the learner in past situations. In this way, the situation will afford the development of courses of interaction exhibiting certain properties that might denote, for example, learning processes that have been cumulative, constructive, self-regulated or reflective, for a learner. IMPLICATIONS FOR A SYSTEM INTELLIGENCE As we have argued, to be consistent with constructivist views of learning a system intelligence should be based on knowledge representation, reasoning, and decision making mechanisms that address the issues that are relevant to constructivist learning, such as the issues discussed in the previous section. These issues indicate that in order to understand learning it is necessary to consider the contexts in which learning takes place, the interactions that happen in these contexts, and the way these interactions are chained over time. Furthermore, in order to facilitate learning it is necessary to consider the affordances of learning situations regarding all these aspects. Concerning the system intelligence, this requires the development of explicit theories that make it possible to formalise the relevant aspects of contexts, interactions, time-extended processes 348 System Intelligence in Constructivist Learning of interaction, and affordances of learning situations, in order to allow reasoning and decision making about these aspects. A context theory will offer means of formalising the content and dynamics of learning situations and the basic aspects of the interactions that develop in these situations, such as the occurrence of learning events or the changes in the content of learning situations caused by learning events. While a context theory allows a system to perceive basic aspects of interaction in situations, it is an interaction theory that helps in interpreting these interactions. Given a formal account of learning situations and of how they change during learning interactions, the role of an interaction theory is to formalise the various ways in which the three aspects of interaction – the situation in which the interaction occurs, the cognitive structures of the learner involved in the interaction, and the nature of the activity that is developed by the learner in the situation – combine to give meaning to learning interactions. The interpretation of single interactions involving context, activity, and cognitive states is a basic step in order to understand time-extended processes of interaction. As a time-extended process, learning depends on the relations that develop over time between aspects of single interactions in situations. Therefore, the role of a theory of time-extended processes of interaction is to formalise the various ways in which interactions relate to one another over time in a course of interaction, to give an account of how process-related qualities of learning processes, such as cumulativeness, constructiveness, self-regulatedness, and reflectiveness, develop in a sequence of interactions with situations. In designing a system intelligence that is attuned to constructivist values, the role of theories of context, interaction, and time-extended process of interaction is to support reasoning about the process of learning in the broad sense that includes the context of learning interactions and the temporal qualities of these interactions, in order to evaluate learning. On the other hand, given an evaluation of learning in these terms, in order to change the environment to facilitate learning, in ways that conform with constructivist views, we need a theory of affordances, which will allow a system to make decisions about the learning opportunities to be provided to a learner whose time-extended process of interaction with the situations of the environment is in a certain state. Therefore, while the state of a learning process is given in precise terms by the theories of context, interaction, and time-extended process of interaction, the utility of a situation for a learner whose learning process is in a certain state, at a certain time, is given by a theory of affordances. In the next sections we describe our approach to the development of these theories illustrating with examples from a simple application in the domain of salad design. Later on, we show how these theories are used to support a system intelligence in a more extended implementation of an intelligent learning environment for the domain of software engineering. FORMALISING THE CONTEXT OF LEARNING The issue of formalising context is becoming central to research in artificial intelligence and related areas (Akman and Surav, 1996). Among the approaches that have been developed, situation theory (Barwise and Perry, 1983; Devlin, 1991) was particularly influential in the development of our approach to formalising the context of learning interactions, although situation calculus (McCarthy and Hayes, 1969), histories (Hayes, 1985) and other related work (Davis, 1990; Reiter, 1991), have also been considered. Below we briefly point to some of the main issues that were relevant to the development of our approach. The basic idea of situation theory is that all sorts of information about the world are organised in terms of situations. According to the theory, agents usually find themselves in, or refer to, situations as structured parts of the world that constitute the context for their behaviour or communication. The main elements of situation theory's ontology are: situations, which represent structured parts of the world; and infons, which represent items of information about the world. Situations are defined intentionally and are related to the infons that hold in the 349 Akhras and Self situation by means of the support relation (situation supports infon). Infons are represented by ordered sets denoted by <<R, a1, ..., an, i>>, where R is a n-place relation; a1, ..., an are the arguments of R; and i is the polarity, which can assume the values 1 or 0, to indicate whether the relation does or does not hold. Some of the entities that can be placed as arguments of the relation are: individuals, relations, spatial locations, temporal locations, and situations. A relevant issue that has not been particularly explored in the formulation of situation theory is the development of explicit mechanisms to represent actions and the changes in situations that may be caused by the occurrence of actions. There is, however, a proposal of an approach in which actions are represented by pairs of sets of infons, in which the first set of infons corresponds to the action precondition, while the second corresponds to the action postcondition (Ohsawa and Nakashima, 1991). A problem associated with reasoning about changes in situations caused by actions is the frame problem (McCarthy and Hayes, 1969), which derives from the fact that although one can apply a temporal representation like the situation calculus to determine what changes follow from the events that happen in the world, one cannot determine the changes that do not follow. As a way of addressing the frame problem in a first-order logic, Davis (1990) has introduced some extensions to situation calculus involving the definition of a set of axioms to assert ways in which particular types of events do not change particular types of states. Generalising this kind of approach, Reiter (1991) has defined a logical theory to specify the effects of actions on states of the world, which includes two kinds of axioms: axioms to specify the conditions for the occurrence of an action (action precondition axioms), and axioms to specify the ways in which actions affect the states of the world (successor state axioms). These axioms, along with a set of general axioms, allow inference of the facts that hold in a new situation after the occurrence of an action. Taking into consideration the issues involved in formalising context, such as the ones discussed above, we have developed an approach for modelling contexts of learning interactions. The formalism is a many-sorted first-order predicate theory for modelling structural information about learning situations as well as temporal information associated with the way situations develop. The entities included in the formalism address the following aspects of contexts of learning interaction: situations, content of situations, dynamics of situations, and situation development. In the next sections we describe this formalism, illustrating with examples taken from an initial application of our approach, which was an intelligent learning environment for the domain of salad design, called SAMPLE (SAlad Making Process-Sensitive Learning Environment), whose goal is to help students learn concepts of salad making. Before we embark on the discussion of the formalism, however, we briefly introduce the characteristics of the contexts for learning interactions provided by SAMPLE. In SAMPLE, the world is populated with tools and salad ingredients. There are seven groups of ingredients, such as, leafy vegetables (e.g. lettuce), or herbs (e.g. parsley). Each of these groups is characterised by a particular set of states through which the ingredient may pass in its preparation before it is added to the salad or dressing. For example, some of the states and transitions of state that characterise ingredients of the leafless vegetables group are: unwashed washed whole chopped In each learning situation provided by SAMPLE there is a set of ingredients available. The tools available for the learner allow basic actions, such as, wash-ingredient, chop-ingredient, add-ingredient-to-salad, or taste-salad. Through these actions the learner can change the states of ingredients, add ingredients to the salad or dressing, dress the salad, or taste the whole preparation. The tasting mechanism gives to the learner feedback from her or his preparation, determining its taste on the basis of the tastes of each individual ingredient that is part of the salad or dressing. 350 System Intelligence in Constructivist Learning Situation Types The notion of context appears in our theory in two ways. First, as contexts for the development of learning interactions, denoted by situation types. Second, as contexts of developed learning interactions, denoted by situations (see later). For example, the description of a set of things that a child can do to play with a doll in the circumstances of a living room in which there is a doll, characterises a situation type, as it is more concerned with possibilities for the development of interaction, and is not located in any particular time. On the other hand, the description of the circumstances of a living room at midnight when Melissa was kissing her doll characterises a particular situation of the above type. Therefore, situation types are intended to denote open worlds for learning interaction, comprising many kinds of entities and holding various possibilities for action. The internal structure of a situation type is defined in terms of two kinds of entities: entities that denote the way things stand in a learning situation - the content of the situation type; and entities that denote the way learners can interact with the other entities of a learning situation - the dynamics of the situation type. To specify that a certain entity x (of content or dynamics) is part of the definition of a situation type s, we use the notation define(x, s). Content of Situation Types To represent the content of situation types we define objects, relations between objects, properties of objects, states of objects and transitions of states, and relations of generalisation and aggregation. Objects are the units of content in situation types and represent the physical or conceptual entities that are part of a learning situation. Objects are represented by n-place predicates, such as in the two examples below taken from SAMPLE. salad ingredient(tomato) To represent physical and conceptual aspects of complex phenomena in learning situations the units provided by objects have to be combined in many different ways, according to the roles that objects perform in relation to each other. This is represented in terms of relations between objects, properties of objects, and states of objects. For example: relation(describe(salad, recipe)) property(ingredient(lettuce), taste(light)) state(ingredient(greens), washed) Besides representing actual states, we might need to represent the states in which objects might be in, and the transitions of states that objects might go through, i.e., their state graphs, which we represent by means of types of states and types of state transitions. For example: state-type(ingredient(tomato), whole) tran-type(ingredient(tomato), whole, sliced) In modelling the content of learning situations, two hierarchical relations that are useful are: generalisations, that characterise is-a relations between sub-class entities and super-class entities, and aggregations, that characterise part-of relations between component entities and aggregate entities. For example: kind(leafy-vegetable, ingredient(watercress)) part(salad, ingredient(cucumber)) Dynamics of Situation Types To represent the dynamics of situation types we define events, preconditions and effects of events, and contexts of events. 351 Akhras and Self Apart from physical and conceptual entities, learning situations have at least one living entity, the learner, and may have several others. These living entities interact with the physical and conceptual entities, and with each other, by means of events. Each of these living entities may have roles attributed to it in the learning situation, and there are a set of events that characterise the actions involved in performing these roles. To define the potential events in a situation type we use the notation event(a, e), where a is an agent and e is an event type. For example: event(learner, wash-ingredient) event(learner, add-ingredient-to-salad) The set of event types that are used to describe the role of an agent in a situation type represents the formal alphabet of that agent. In the above example, the role defined for the learner in the situation type includes doing things like washing ingredients, and adding them to a salad. The conditions of activation of an event and the changes in a learning situation caused by the occurrence of an event are stated in the preconditions and effects defined for the event type, which are denoted by pre(e, x, pa), and effect(e, x, pa), where e is an event type, x is any content entity and pa is the participation of x in the precondition or effect, which can assume the values 1 or 0 to indicate whether x must hold or not to satisfy the precondition, or whether x will hold or not in the effect. For instance: pre(e, x, 1) means that the precondition for e is x pre(e, x, 0) means that the precondition for e is not(x) Preconditions and effects are particularly important to capture the circumstances involved in a learning event which are essential for the definition of our formal account of interaction. Events are sometimes associated in particular ways to other content of a situation type, which may refer to the background for the event, the sociocultural aspects related to the event, the authentic setting of the event, and so on. To capture this sort of relation we introduce the notion of context of an event, which we denote by context(e, x), where e is an event type and x is a content entity that characterises a context for events of type e. As an example, suppose that among the entities that are part of the content of learning situations in SAMPLE, there is this book: book("Well balanced salads"), which may characterise a context for events of type addingredient-to-salad. This is represented as: context(add-ingredient-to-salad, book("Well balanced salads")) Situation Development Interactions develop in a situation type by the occurrence of events and give rise to situations. Although situation types are independent of time, a situation is temporally located and denotes the state of a situation type at a certain time. Situations are, thus, defined by the pair: (situation type, time) Events occur in situation types at certain times, which is the same as saying that events occur in situations. To denote the occurrence of events in situations we introduce the notation occurs(e, a, s, t), where e is an event type, a is an agent, and (s, t) is the situation in which the event occurs. For example: occurs(wash-ingredient, learner, salad-lab-a, 6) To denote the content entities that hold in situations we introduce the notation in(x, s, t), where x is any content entity, and (s, t) is the situation in which the content entity is present. For example, the following content entities may hold in the situations before and after the occurrence of the event above, if the ingredient washed is a lettuce: in(state(ingredient(lettuce), unwashed), salad-lab-a, 6) in(state(ingredient(lettuce), washed), salad-lab-a, 7) 352 System Intelligence in Constructivist Learning In order to address the problem of tracking the changes that occur in situations through interactions, which is an instance of the frame problem, we have defined three kinds of axioms. A set of axioms of situation development, called effect axioms, specify the content entities that must hold at the end of an event that occurs if the content entities that are preconditions for the event hold at the beginning of that event. Table 1 presents the first two of these axioms. Other two axioms (SD3 and SD4) are variations of SD-1 and SD-2 in which the value of the participation of content entities in the definitions of effects is 0. Table 1. Axioms of situation development _____________________________________________________________________ Effect axioms SD-1: occurs(e, a, s, t) define(pre(e, x, 1), s) define(effect(e, y, 1), s) in(x, s, t) in(y, s, t+1) occurs(e, a, s, t) define(pre(e, x, 0), s) define(effect(e, y, 1), s) in(x, s, t) in(y, s, t+1) _____________________________________________________________________ SD-2: The effect axioms allow us to infer the changes in content entities caused by the occurrence of events in situations. However, they do not solve the general problem of determining the content entities that hold in a certain situation. This is determined in our theory by taking into consideration the histories of participation of content entities in the preconditions or effects of events that have occurred, as described below. As well as content entities that can hold in situations, actual preconditions and effects of events can also hold in situations. This is represented using the notation in(y, s, t), as before, with y taking the form of pre(e, x, pa) or effect(e, x, pa) to denote the actual preconditions and effects that hold at the beginning and at the end of an event e, with x being an instantiated content entity and pa assuming 0 or 1. For example, some of the actual preconditions and effects that hold when the event wash-ingredient occurs, are: in(pre(wash-ingredient, state(ingredient(lettuce), unwashed), 1), salad-lab-a, 6) in(effect(wash-ingredient, state(ingredient(lettuce), washed), 1), salad-lab-a, 7) These actual preconditions and effects of events characterise points in the histories of participation of content entities in the preconditions and effects of events that occur. These histories are used in determining the content entities that hold in situations. In addition, we have described earlier that the initial state of a situation type is given by the content of the situation type defined using the formula define(x, s). Now, when the interaction in a situation type s begins, events that occur give rise to histories of participation of content entities in preconditions and effects of events, which characterise changes in the initial state of s. Therefore, to determine whether certain content entities do or do not hold in a situation we must take into consideration this initial definition of the situation type as well as the changes of content caused in situations by the occurrence of events, characterised by these histories. To allow this kind of inference we have formulated three further axioms of situation development: SD-5 and SD-6, which refer to points of histories of content entities that hold in situations, and SD-7, which allows to infer the content entities that hold in situations. These axioms are shown in Table 2. As interactions progress over time and situations of a single or various types are developed, sequences of situations are formed, giving rise to the development of courses of interaction. A course of interaction is defined in our theory by a sequence of situations and is denoted by course(s1, t1, ..., sn, tn), where (s1, t1) and (sn, tn) are any two situations, which can possibly be of the same type, and for which n>=2 and tn>t1. A particular case of a course of interaction is the course of two situations course(s1, t1, s2, t2), which we will be using in the rest of this paper. 353 Akhras and Self Table 2. Further axioms of situation development _____________________________________________________________________ History axioms SD-5: SD-6: define(pre(e, x, pa), s) occurs(e, a, s, t) in(pre(e, x, pa), s, t) define(effect(e, x, pa), s) occurs(e, a, s, t) in(effect(e, x, pa), s, t+1) State axiom [in(pre(e, x, 1), s, t) in(effect(e, x, 1), s, t)] t1)[t1<t in(x, s, t1)] t2)[t1<t2<t in(effect(e, x, 0), s, t2)] [define(x, s) t1)[t1<t in(effect(e, x, 0), s, t1)] in(x, s, t) _____________________________________________________________________ SD-7: Note that according to this definition, courses of interaction are not necessarily contiguous, as there may be other situations located between (s1, t1) and (s2, t2). Therefore, courses of interaction can overlap in many ways, and a situation can appear in more than one course of interaction. As situations are a way of preserving the context of events that occur, courses of interaction are a way of preserving the history of the interaction, which is essential to the analysis of properties of time-extended processes of interaction. UNITS OF ANALYSIS OF LEARNING INTERACTIONS Research on interaction in artificial intelligence is recent and has aimed at the development of computational theories of agents' involvements in their environments, with two main purposes: to guide the analysis of living agents and the design of artificial ones. The central point of the work on computational theories of interaction and agency, as reported by Agre (1995), is to take an interactional perspective on understanding the behaviour of agents in their environments. Instead of units of analysis based on the agent's cognitive process, the focus is on units of analysis that refer to interactions, whose definition requires research focused on the discovery of structures in the world and of properties of interactions. To illustrate, Agre (1995) considers a controller (the agent) of an oil refinery (the environment), with the general task of the agent being the adjustment of certain devices in its environment so that a desired flow of oil is sustained within the refinery. Given a proposed design for this controller, how can we know whether it will work? The answer, it is argued, cannot rely only on an analysis of the controller itself, and obviously, nor on an analysis of the plant in isolation. Instead, it is crucial to analyse how the controller will interact with the plant. Therefore, in order to understand an agent’s interaction with its environment this approach focuses on reasoning to recognise structures in the relationships among the properties of agents, environments, and forms of interaction between them (which may not have an internal state in the agent’s mind), rather than on the more classical AI approach of reasoning to recognise plans that can be attributed to an agent. Concerning human learners interacting in their learning environments, the interactional perspective that is necessary to interpret learning phenomena, as we have discussed, requires that we take the three aspects that characterise a learning interaction - context, activity, and cognitive structures, and look for regularities in the ways these three aspects relate to each other in interactions that are developed in learning processes. In our model of learning situations, we have defined contexts of learning interactions as types of situations and introduced a set of formal entities to denote the content of these contexts, and the activities that can be developed on them, which characterise the dynamics of these contexts. In addition, we have defined a set of formal entities to denote aspects of the interactions that are developed in these contexts when the potential activities defined in the dynamics of a situation type actually occur. These three sets of formal entities: entities of content, entities of dynamics, and entities of situation development, are the basic elements from 354 System Intelligence in Constructivist Learning which a set of types of regularities of interaction in learning situations, which we call patterns of interaction, can be formally defined. Therefore, the formalism introduced earlier for modelling structural information about learning situations as well as temporal information associated with the way situations develop, is now extended to include patterns of interaction, which are formal entities that model higherorder regularities of interaction in learning situations. Exploring the connections that are formed as learners interact in situations, among aspects of interaction such as the context of activity, the nature of activity, and the cognitive states that learners bring to activity, we have defined three types of patterns of interaction: patterns that relate learner's actions to the situations in which they happen, patterns that relate learner's cognitive states to the situations in which they hold, and patterns that capture relations between situations. In the next sections we discuss these patterns and present some of their formal definitions, illustrating with examples taken from SAMPLE. Patterns of Learner's Actions in Situations Situations develop by the occurrence of events which may affect or be affected in various ways by the content of the situation. Therefore, the nature of a learner's action in a situation is given by the way it affects or is affected by the content of the situation. The different ways in which this happens characterise different patterns of learner's actions in situations. Following the discussion about the role of activity in learning, our definitions of these patterns intend to capture the various ways in which meaning is constructed from acting in situations. Some of these definitions, which correspond to learner's actions of utilising, generating, and accessing entities in situations, are: Definition (Utilising entities in situations) A learner a utilises an entity x through an event e in a situation (s, t), iff the event e, defined in situation type s as part of the alphabet of the learner a, occurs in (s, t), and x is a precondition of e that holds in (s, t), i. e. the participation of x in the precondition is 1. define(event(a, e), s) occurs(e, a, s, t) in(pre(e, x, 1), s, t) utilises(a, x, e, s, t) For example, suppose that we have the situation type salad-lab-a in which the following entities are defined: define(event(learner, wash-ingredient), salad-lab-a) (1) define(pre(wash-ingredient, ingredient(X), 1), salad-lab-a) define(pre(wash-ingredient, state(ingredient(X), unwashed), 1), salad-lab-a) define(effect(wash-ingredient, state(ingredient(X), washed), 1), salad-lab-a) And suppose that in situation (salad-lab-a, 8), the learner washes a lettuce, and the following entities hold: in(ingredient(lettuce), salad-lab-a, 8) in(state(ingredient(lettuce), unwashed), salad-lab-a, 8) occurs(wash-ingredient, learner, salad-lab-a, 8) (2) in(pre(wash-ingredient, ingredient(lettuce), 1), salad-lab-a, 8) (3) in(pre(wash-ingredient, state(ingredient(lettuce), unwashed), 1), salad-lab-a, 8) (4) Therefore, according to the definition of the pattern utilises and according to the expressions (1), (2) and (3) above, we say that the learner utilises the ingredient lettuce by washing it in the situation (salad-lab-a, 8), which is the same as saying that the following pattern holds: utilises(learner, ingredient(lettuce), wash-ingredient, salad-lab-a, 8) Additionally, note that according to the expression (4) above, the learner also utilises in the same event the notion of an ingredient lettuce being in a state of unwashed. In SAMPLE, some of the main things that a learner can utilise are: all sorts of ingredients, salad and dressing, the 355 Akhras and Self notions of ingredients being in some sorts of states, and the notions of some sorts of transitions of ingredients' states. Definition (Generating entities in situations) A learner a generates an entity x through an event e in a situation (s, t), iff the event e, defined in situation type s as part of the alphabet of the learner a, occurs in (s, t), and x is an effect of e that holds in (s, t+1). define(event(a, e), s) occurs(e, a, s, t) in(effect(e, x, 1), s, t+1) generates(a, x, e, s, t) In the situation (salad-lab-a, 8) of the previous example, after the learner has washed the ingredient lettuce, the following entities hold in situation (salad-lab-a, 9): in(effect(wash-ingredient, state(ingredient(lettuce), washed), 1), salad-lab-a, 9) in(state(ingredient(lettuce), washed), salad-lab-a, 9) (5) Therefore, according to the definition of the pattern generates and according to the expression (5) above, we say that the learner generates the notion of an ingredient lettuce being in a state of washed, by washing it in the situation (salad-lab-a, 8), or, more formally: generates(learner, state(ingredient(lettuce), washed), wash-ingredient, salad-lab-a, 8) In general, in interaction with SAMPLE a learner can generate: the notions of ingredients being in some sorts of states, the notions of an ingredient being part of a salad or dressing, and the taste of a salad. Here we can see how some limitations may be identified in learning environments such as SAMPLE. Ingredients, which are the main building blocks in the salad preparation world for the learner, can be utilised but cannot be generated. This is because in many learning environments (computational or not) some things happen to be ready for the learner, requiring no construction. This has strong implications for learning environments that intend to achieve a higher level of constructiveness, as we will discuss later. Definition (Accessing entities in situations) A learner a accesses an entity x through an event e in a situation (s, t), iff the learner a utilises the entity x in situation (s, t), and x is a precondition of e that holds in (s, t), and the learner does not generate any entity in the same event. utilises(a, x, e, s, t) in(pre(e, x, 1), s, t)y)generates(a, y, e, s, t) accesses(a, x, e, s, t) In situations of SAMPLE, the learner can access: information in books or archives, the contents of the salad or dressing being prepared, and characteristics of ingredients such as their tastes or their current states. Patterns of Learner's Cognitive States in Situations According to constructivist views of learning, cognitive structures develop from acting in situations. Therefore, some relevant cognitive structures may be developed from actions of utilising, generating, or accessing entities in situations. These cognitive structures influence in many ways the learner's view of the content and dynamics of subsequent situations. Following the discussion about the role of cognitive structures in learning, our definitions of these patterns are intended to capture the various ways in which entities of a situation are related to the learner's previously formed cognitive structures. Some of these definitions, which correspond to the learner's cognitive states in which entities of situations are new or old, are: Definition (New content entities in situations) A content entity x is new for a learner a in a situation (s, t), iff the entity x holds in situation (s, t), and x has neither been utilised nor generated by the learner a in any situation previous to (s, t). in(x, s, t) (ei, si, ti)[ti<t utilises(a, x, ei, si, ti)] (ej, sj, tj)[tj<t-1 generates(a, x, ej, sj, tj)] new(a, x, s, t) 356 System Intelligence in Constructivist Learning For an example, consider the situation (salad-lab-a, 8) of the previous example, in which the following entity holds: in(state(ingredient(lettuce), unwashed), salad-lab-a, 8) Assuming that the entity state(ingredient(lettuce), unwashed) has neither been utilised nor generated by the learner before, then according to the definition of the pattern new for content entities, we say that the notion of an ingredient lettuce being in a state of unwashed is new for the learner in the situation (salad-lab-a, 8), which means that the following pattern holds: new(learner, state(ingredient(lettuce), unwashed), salad-lab-a, 8) Definition (Old content entities in situations) A content entity x is old for a learner a in a situation (s, t), iff the entity x holds in situation (s, t), and x has been utilised or generated by the learner a in some situation previous to (s, t). in(x, s, t) [ (ei, si, ti)[ti<t utilises(a, x, ei, si, ti)] (ej, sj, tj)[tj<t-1 generates(a, x, ej, sj, tj)] ] old(a, x, s, t) Note that old is different from not new because in both cases the entity that is new or old must be present in the situation, as denoted by the primitive in(x, s, t). This derives from the fact that these patterns refer to cognitive states in relation to situations. Therefore, if an entity is not new and is not present in a situation it does not characterise the kind of old that we are capturing in these particular patterns, although in the common sense of the word it would be old. Patterns of Relations Between Situations A characteristic of the constructivist view of learning that we have discussed is that learning occurs in situations that correspond to real life contexts and, therefore, requires multiple types of situations where varied perspectives are portrayed and learners can explore various aspects of a domain. As learners go from situation to situation, interacting in this kind of environment, they are likely to connect through experience knowledge of different kinds, and these experiences and the connections that derive from them will be influenced by the relations that exist between entities of situations. The different ways in which entities of one situation may be related to entities of another situation characterise different patterns of relations between situations. Therefore, our definitions of these patterns are intended to capture the various ways in which entities of one situation are related to entities of another situation. The patterns that we have defined are of two kinds: patterns of relations in which situations share some characteristic, and patterns of relations in which a situation has an additional, but related, characteristic with respect to another situation. The specific patterns of each of these kinds correspond to the different ways in which situations can share characteristics or have additional characteristics with respect to other situations. The definition of one of these patterns is shown below (other similar definitions are given in Akhras(1997)). Definition (Sharing content entities) Two situations (s1, t1) and (s2, t2) share a content entity x, iff x holds in situation (s1, t1) and in situation (s2, t2). in(x, s1, t1) in(x, s2, t2) share(s1, t1, s2, t2, x) UNITS OF ANALYSIS OF TIME-EXTENDED LEARNING PROCESSES Courses of interaction are formed from sequences of situations that develop by the occurrence of events when learners are engaged in interaction with the situation types of their learning environments. As we have discussed earlier, the focus on the situations that develop when events occur, rather than on the events alone, to account for the progression of the interaction between learner and environment, is a way of preserving the context in which the events take 357 Akhras and Self place. The preservation of the context in understanding learning phenomena was a major point in the development of our formal account of interaction in the previous section, where a set of patterns of interaction were formally defined in terms of more basic entities. Now, a second major point of our theory is the preservation of the history of the interaction, which encompasses the process and structure by which aspects of interactions developed in different situations are connected to one another during a course of interaction. The focus on sequences of situations to account for courses of interaction is a way of preserving the history of the interaction. Histories of interaction when preserved will embed information about particular ways in which courses of interaction develop, according to the particular patterns of interaction that hold in the situations of the course of interaction. These particular ways in which courses of interaction develop will characterise regularities of a higher order than the patterns of interaction as they will relate patterns of interaction that hold in different situations. To model these regularities, our theory (that already encompasses formal entities to denote the content and dynamics of situation types, aspects of situation development and patterns of interaction) is now extended to include properties of courses of interaction, which are formal entities that denote regularities of time-extended learning processes. In order to model some of the process-related notions that are addressed by constructivist learning approaches, we have defined four types of properties of courses of interaction: cumulativeness, constructiveness, self-regulatedness, and reflectiveness. These properties are defined in terms of patterns of interaction and of entities of situation type and of situation development. In the next sections we discuss these properties and present their formal definitions (variations of these definitions are given in Akhras (1997)), illustrating with examples taken from SAMPLE. Although the approach can be applied to any kind of course of interaction, the properties that we have defined are based on courses of interaction involving only two situations. Property of Cumulativeness Cognitive conceptions of learning stress that learning is cumulative in nature (Shuell, 1986). Nothing has meaning or is learned in isolation. Instead, prior knowledge, and consequently, previous learning experiences, influence and relate to new learning in many ways. Repetition of similar experiences in different contexts and involving different ways of looking at the experiences may enable access to prior knowledge and the exploitation of similarities and differences between the current and the previous experiences. Ultimately, this leads to a cumulative process in which new meanings are added to elements of previous experiences and current experiences are interpreted in the light of previous ones. Although cumulativeness alone may not be an indicator of learning, it is part of a learning process. In our theory, cumulativeness refers to the property that a course of interaction exhibits when entities experienced by the learner in one situation are in some way revisited in a later situation of the course of interaction. A particular way in which a course of interaction can be cumulative is through a shared entity, which happens when the same entity is experienced in the two situations of a course of interaction. Other ways may involve experiencing entities in the two situations that are not the same but are in some ways related. Below we present the definition of cumulativeness from a shared entity. Definition (Cumulative with respect to a shared content entity) A course of interaction course(s1, t1, s2, t2) is cumulative with respect to a content entity x for a learner a, if situations (s1, t1) and (s2, t2) share the entity x, and the learner a utilises x in (s1, t1) or generates it in (s1, t1-1), and further utilises x in (s2, t2) or generates it in (s2, t2-1). share(s1, t1, s2, t2, x) [utilises(a, x, e1, s1, t1) generates(a, x, e1, s1, t1-1)] [utilises(a, x, e2, s2, t2) generates(a, x, e2, s2, t2-1)] t2>t1 cumulative(course(s1, t1, s2, t2), a, x) 358 System Intelligence in Constructivist Learning For example, suppose that the following patterns of interaction hold in the situations below. share(salad-lab-a, 8, salad-lab-a, 11, ingredient(lettuce)) utilises(learner, ingredient(lettuce), wash-ingredient, salad-lab-a, 8) utilises(learner, ingredient(lettuce), add-ingredient-to-salad, salad-lab-a, 11) Then, according to the definition above, we say that the course of interaction course(saladlab-a, 8, salad-lab-a, 11) is cumulative for the learner with respect to the ingredient lettuce, which means that the following property holds: cumulative(course(salad-lab-a, 8, salad-lab-a, 11), learner, ingredient(lettuce)) Property of Constructiveness In essence, to construct new knowledge involves relating one's existent knowledge to new experiences in meaningful ways. In this process, not only is the knowledge associated with the new experience constructed but also the learner's existent knowledge is sometimes reinterpreted in the light of the new experience. According to Shuell (1992), learning is constructive in the sense that the new information that is perceived and interpreted by the learner in a unique way must be elaborated and related to other information in order that it can be learned. Therefore, an essential feature of learning processes is the integration of aspects of new learning experiences with the learner's existent knowledge. As the learner interacts in situations, information from several sources, including previous experiences, must be elaborated and combined in meaningful ways, so that the new information that is generated and interpreted by the learner can be related to the learner's existent knowledge, which may also be re-interpreted in the light of the new experience. Ultimately, this leads to a constructive process in which new knowledge is generated and related to elements of previous experiences. In our theory, constructiveness refers to the property that a course of interaction exhibits when entities experienced by the learner in one situation are in some way related to new entities that the learner generates in a later situation of the course of interaction. A particular way in which a course of interaction can be constructive is from an event, which involves experiencing an entity that is old for the learner, in one situation of a course of interaction, and further generating a new entity, through an event that utilises the old entity, in another situation of the course of interaction, which will then be constructive with respect to the new entity. This definition is presented below. Other ways may involve generating a new entity that is connected to the old entity in several other ways. Definition (Constructive with respect to a content entity from an event) A course of interaction course(s1, t1, s2, t2) is constructive with respect to a content entity x for a learner a, if the learner a utilises an entity xo in (s1, t1) or generates it in (s1, t1-1), and further utilises xo in (s2, t2-1) which is old for the learner in this situation, to generate in the same event an entity x which is new for the learner in situation (s2, t2). [utilises(a, xo, e1, s1, t1) generates(a, xo, e1, s1, t1-1)] utilises(a, xo, e, s2, t2-1) old(a, xo, s2, t2-1) generates(a, x, e, s2, t2-1) new(a, x, s2, t2) t2>t1 constructive(course(s1, t1, s2, t2), a, x) For example, suppose that the following patterns of interaction hold in the situations below. utilises(learner, ingredient(lettuce), chop-ingredient, salad-lab-a, 22) utilises(learner, ingredient(lettuce), add-ingredient-to-salad, salad-lab-a, 28) old(learner, ingredient(lettuce), salad-lab-a, 28) generates(learner, part(salad, ingredient(lettuce)), add-ingredient-to-salad, salad-lab-a, 28) new(learner, part(salad, ingredient(lettuce)), salad-lab-a, 29) Then, according to the definition above, we say that the course of interaction course(saladlab-a, 22, salad-lab-a, 29) is constructive for the learner with respect to the notion of an 359 Akhras and Self ingredient lettuce being part of a salad, which is the same as saying that the following property holds: constructive(course(salad-lab-a, 22, salad-lab-a, 29), learner, part(salad, ingredient(lettuce))) Note that, although learning comes from acting, our definitions of cumulativeness and constructiveness take not just the patterns utilises or generates (i.e. not just acting), but also other patterns of interaction, relating acting to aspects of contexts and cognitive states involved in action. Property of Self-Regulatedness As learners interact in situations they have to make decisions about what actions to take in order to attain their goals or even to help in defining their goals. This requires an awareness of how they are progressing in their learning experiences and an ability to regulate their involvement in these experiences. In a constructivist view of learning, the activities that bring this awareness and help in regulating the learner's actions are performed by the learner, and involve learners regulating their actions based on several kinds of information that they obtain from their interactions in situations. Ultimately, this leads to a self-regulated process in which aspects of the learning experience are analysed and used to drive the learner's actions. In our theory, self-regulatedness refers to the property that a course of interaction exhibits when a learner's action performed in one situation is in some way evaluated by the learner in another situation of the course of interaction, and this evaluation is taken into consideration to guide the next learner's actions or change the effects of previous actions. The learner's actions correspond to events that are part of the learner's alphabet and occur in situations. These events have an associated context which represents the information that is relevant to evaluating the corresponding actions. This information can be defined as part of the situation type and accessed when needed or be generated by the learner in a dynamic evaluation. The different ways in which this information is produced and used characterise different ways in which a course of interaction can be self-regulated for a learner. Among the many different ways in which a course of interaction can be self-regulated, we have identified two main classes of self-regulatedness: self-regulatedness from access and selfregulatedness from generation, which involve acting in one situation of a course of interaction, and accessing (or generating) information that helps in evaluating that action, in another situation of the course of interaction. The information accessed (or generated) is an evaluation context for that action, and the course of interaction will then be self-regulated with respect to the result of the action from the point of view of the evaluation context accessed (or generated). Below we present a definition of self-regulatedness from access. Definition (Self-regulated with respect to a content entity, accessing the context before the event) A course of interaction course(s1, t1, s2, t2) is self-regulated with respect to generating a content entity x through an event e and accessing a context c before the event, for a learner a, if the learner a accesses the entity c that is an evaluation context for the event e, in (s1, t1), and further generates the entity x in (s2, t2) through the event e, with t2-t1 being the time gap between accessing the evaluation context and performing the related action. accesses(a, c, e1, s1, t1) in(context(e, c), s1, t1) generates(a, x, e, s2, t2) t2>t1 t2-t1=time gap self-regulated(course(s1, t1, s2, t2), a, x, e, c) For an example, suppose that the following context is defined in situation type salad-lab-a for the event type add-ingredient-to-salad: context(add-ingredient-to-salad, book("Well balanced salads")) In addition, suppose that the following entity of situation development and patterns of interaction hold in the situations below. 360 System Intelligence in Constructivist Learning in(context(add-ingredient-to-salad, book("Well balanced salads")), salad-lab-a, 26) accesses(learner, book("Well balanced salads"), view-book, salad-lab-a, 26) generates(learner, part(salad, ingredient(lettuce)), add-ingredient-to-salad, salad-lab-a, 28) Then, according to the definition above, we say that the course of interaction course(saladlab-a, 26, salad-lab-a, 28) is self-regulated for the learner with respect to adding an ingredient lettuce to a salad, in the context provided by the book "Well balanced salads", which means that the following property holds: self-regulated(course(salad-lab-a, 26, salad-lab-a, 28), learner, part(salad, ingredient(lettuce)), add-ingredient-to-salad, book("Well balanced salads")) Property of Reflectiveness When learners interact in situations and develop their own activities, for instance, to solve a problem or to construct an artefact, reflection involves them being aware of the process by which they are developing these activities, and to take this process as the object of their thinking. According to Dewey (1933), quoted in (Ertmer and Newby, 1996), reflection involves "reconstruction or reorganization of experience which adds to the meaning of experience and which increases ability to direct the course of subsequent experience" (p.76). This reflective activity allows the learners to focus on the process of interacting and learning in situations, rather than on the product. The first main activity required for reflection about a particular experience is the representation of that particular experience, or, as von Glasersfeld (1995) would prefer: the representation of the experience, as it involves the learner presenting again, or replaying, to herself a past experience. In this process, the learner has to recollect what has taken place in the experience and replay the events that have happened noticing everything that might be relevant (Boud, Keogh and Walker, 1985). The second main activity, which comes after this recollection of information about an experience, is the evaluation. Based on this information the learner evaluates the experience, focusing primarily on the process by which the course of interaction has developed. This allows the learner to determine how effective her overall process was in achieving her goals and to determine the extent of her achievements (Ertmer and Newby, 1996; Simons, 1993). Ultimately, this leads to a reflective process in which the learners assess their learning experiences focusing primarily on the overall process by which their interactions in situations have developed. In our theory, reflectiveness refers to the property that a course of interaction exhibits when aspects of the learner's process of interaction in some situations are the objects of reflective activities carried out by the learner in later situations of the course of interaction. Basic aspects of the process of interaction are represented by the entities that denote the occurrence of events and the presence of entities in situations. These entities are basic process entities and will constitute the basic elements of other types of process entities defined in types of situations. For example, a trace, which is a kind of process entity, can be generated as a sequence of event occurrences, and its generation can be defined as part of the dynamics of situation types. Similarly to the case of self-regulation, we have identified two main classes of reflectiveness: reflectiveness from access and reflectiveness from generation, which involve acting in a set of situations of a course of interaction, and further accessing (or generating) an entity that contains information about the process of interaction in those situations, in another situation of the course of interaction, which will then be reflective with respect to the entity accessed (or generated). Below we present a definition of reflectiveness from access. In this definition we use the term process entity to denote particular kinds of content entities that contain information about aspects of the process of interaction. Formally, a process entity is a content entity that is related in some ways to an entity of situation development, such as occurs( ) or in( ). Therefore, the two main types of process entities are: process entities derived from collecting occurrences of events in situations, and process entities derived from collecting states of situations or presences of entities in situations. 361 Akhras and Self Definition (Reflective from access with respect to occurrences) A course of interaction course(s1, t1, s2, t2) is reflective with respect to a process entity x that collect occurrences of events in situations, for a learner a, if the process entity x is generated or updated in (s1, t1) and further accessed by the learner in (s2, t2). generates(a, part(x, occurs(e1, a, s1, t1)), e1, s1, t1) accesses(a, x, e2, s2, t2) t2>t1 reflective(course(s1, t1, s2, t2), a, x) For an example, suppose that the following process entity can be accessed in situation type salad-lab-a to show the sequence of events which occurred in the process of interaction with that situation type, where t1 is the time of the first occurrence and tn is the time of the last occurrence. trace-salad(salad-lab-a, t1, tn) Accessing this entity after developing the events that are collected in the trace will characterise a reflective course of interaction. The properties that we have discussed and defined above characterise regularities that may happen in courses of interaction which denote particular ways in which learning processes can be cumulative, constructive, self-regulated or reflective. Following the same approach we may, of course, define variations and refinements of these definitions and may seek to define further properties. Our definitions could also be extended to adopt more complex models of these properties, such as Winne and Hadwin's (1998) model of self-regulation, or the model of selfregulation and reflection described by Ertmer and Newby (1996). AFFORDANCES OF LEARNING SITUATIONS According to the theory that we have described in the previous sections, courses of interaction develop properties such as cumulativeness, constructiveness, self-regulatedness and reflectiveness, when learning events that occur in the situations involved in these courses of interaction lead to certain patterns of interaction that are relevant for the development of the particular properties. Therefore, after a sequence of learning events, several patterns of interaction hold in the situations where the learner had been interacting, and a set of properties of courses of interaction hold as a consequence of these patterns. These patterns of interaction and properties of courses of interaction that hold throughout the situations of the learning process up to a certain time characterise the state of the learning process at that time. Now, suppose that a learner has interacted for some time with an environment, in several situation types, and is about to engage in interaction with a further situation type. Each event available for the learner in this new situation type characterises a possibility for interaction. However, not all events characterise the same possibilities for learning, i.e., not all events lead to the development of courses of interaction that exhibit the same properties. Whether or not possibilities for interactions entail possibilities for properties of courses of interaction will depend on the state of the learning process. Therefore, according to the characteristics of the events defined as part of the dynamics of a given situation type, and according to the previous history of the learner's interaction with other situation types, interaction in this new situation type may lead to the development of particular patterns of interaction, and consequently, allow certain properties of courses of interaction to hold. If a system can know in advance the possibilities offered by a situation type for the development of these properties for a learning process that is in a certain state, then the system can use this information to support its decision making concerning the kinds of situation types that will be made available for the learner in the environment's space of interaction at a particular time. In order to allow an intelligent system to consider these possibilities in advance, we have developed a formal account of possibilities for interactions and for time-extended processes of interaction that exhibits particular characteristics, based on the notion of affordance introduced 362 System Intelligence in Constructivist Learning by Gibson (1977). Formalising affordances will make possible for an intelligent learning environment to reason about the features of situation types that afford the desired patterns of interaction and properties of courses of interaction to be developed in the following events of a learning process that is in a certain state. Concerning interactions, a situation type may afford to a learner the development of particular patterns of interaction. For example, a situation type s in which the following definitions hold: define(event(a, e), s) define(effect(e, x, pa), s) affords to the learner a generating the entity x through the event e (according to the definition of the pattern generates). For another example: if an entity of a situation type is new to a learner and there are ways in the situation type in which the learner can utilise this entity, which then would become old for the learner (according to the definition of the pattern old), then we can say that an affordance of this situation type to that learner is the possibility of developing the pattern of interaction in which the referred entity is old for the learner. Of course, after this event happens, i.e. the utilisation of the entity and it becoming old, the affordance will no longer be there, although the learner may still be interacting with the same situation type. In addition to what a situation type can afford to a learner in an interaction, namely the development of patterns of interaction such as the ones that we have defined earlier, a situation type can also afford to a learner the development of courses of interaction that exhibit particular properties. For example, suppose that the following pattern of interaction holds in a situation (s1, t1) in which the learner has interacted: utilises(a, x, e1, s1, t1) and suppose that the same definitions of the previous example hold for the situation type s. Then, the occurrence of the event e at time t, would indicate the development of the following patterns of interaction: share(s1, t1, s, t, x) generates(a, x, e, s, t) and, if we take into consideration the course of interaction from situation (s1, t1) to the now developed situation (s, t), the holding of these patterns, according to the definition of cumulativeness, would indicate the development of a course of interaction course(s1, t1, s, t+1) that is cumulative with respect to the entity x for the learner a. Therefore, we can say that the situation type s affords to the learner a, at a certain time greater than t1, the development of a course of interaction that is cumulative with respect to the entity x. Therefore, concerning time-extended processes, a situation type may afford to a learner the development of courses of interaction that possess particular properties such as the ones that we have defined in the previous section. Note that the way in which the affordance is relative to the learner refers to relativeness to the whole process that the learner has gone through, and not only to individual characteristics of the learner. In the example above, if the learner has not utilised the entity x in any previous situation, the situation type s would not be able to afford cumulativeness with respect to that entity at this time. In our theory, we have identified two types of affordances of situation types: affordances for patterns of interaction, which represent possibilities in situation types for the development of patterns of learner's actions in situations, learner's cognitive states in situations, or relations between situations, and affordances for properties of courses of interaction, which represent possibilities in situation types for the development of courses of interaction that are cumulative, constructive, self-regulated, or reflective. Below we present some of these definitions (a more extended set of definitions of these two kinds of affordances is given in Akhras(1997)). 363 Akhras and Self Affordance for Patterns of Learner's Actions in Situations In order that a situation type can afford to a learner the development of patterns of learner's actions in situations, the situation type must contain types of events with the kinds of preconditions and effects that allow the learner to utilise, generate, or access entities of the content of the situation type, according to the definitions of these patterns presented before. Below is the definition of one of these affordances. Definition (Affords generating entities in situations) A situation type s affords to a learner a generating an entity x through an event e, iff the event e is defined in situation type s as part of the alphabet of the learner a, and x is an effect of e defined in s. define(event(a, e), s) define(effect(e, x, pa), s) affords(s, generates, a, x, e) For an example, consider the effect of the event type slice-ingredient as defined in situation type salad-lab-a: define(event(learner, slice-ingredient), salad-lab-a) define(effect(slice-ingredient, state(ingredient(X), sliced), 1), salad-lab-a) Then, according to the definition above, we say that the situation type salad-lab-a affords to the learner generating the notion of an ingredient being in a state of sliced through slicing it, which means that the following affordance holds: affords(salad-lab-a, generates, learner, state(ingredient(X), sliced), slice-ingredient) Affordance for Patterns of Learner's Cognitive States in Situations To afford to a learner the development of patterns of learner's cognitive states in situations, a situation type must contain the kind of content and dynamics that allow a learner who is in a certain cognitive state to develop the patterns in which entities of the situation type become new or old to the learner, according to the definitions of these patterns presented before. Below we present the definition of the affordance for the pattern new, in which these conditions are formally stated. Definition (Affords new content entities in situations) A situation type s affords to a learner a an entity x being new at the current time tc or after, iff the entity x is defined or can be generated in the situation type s, and x has not been utilised nor generated by the learner a in any situation previous to (s, tc). [define(x, s) affords(s, generates, al, x, el)] (ei, si, ti)[ti<tc utilises(a, x, ei, si, ti)] (ej, sj, tj)[tj<tc-1 generates(a, x, ej, sj, tj)] affords(s, new, a, x, tc) For example, if the state of an ingredient cabbage is defined in a situation type as unwashed, and no pattern of utilising or generating this ingredient state holds in previous situations, then, according to the definition above, we say the situation type affords to the learner, at the current time or after, the notion of an ingredient cabbage in a state of unwashed being new, which is the same as saying that the following affordance holds: affords(salad-lab-a, new, learner, state(ingredient(cabbage), unwashed), tc) Affordance for Cumulativeness In order that a situation type can afford to a learner the development of courses of interaction that are cumulative, the situation type must afford to the learner the development of the kinds of patterns of interaction that will lead to particular kinds of cumulativeness being developed, according to the definitions of cumulativeness presented before. Below we present the definition of one of these affordances. 364 System Intelligence in Constructivist Learning Definition (Affords cumulativeness with respect to a content entity) A situation type s affords to a learner a the development of a course of interaction from situation (si, ti) that is cumulative for the learner with respect to a content entity x, if the learner utilises x in (si, ti) or generates it in (si, ti-1) and the situation type s affords to the learner a situation that shares the entity x with situation (si, ti) and also affords to the learner utilising or generating x. utilises(a, x, ei, si, ti) generates(a, x, ei, si, ti-1)] affords(s, share, si, ti, x) [affords(s, utilises, a, x, e) affords(s, generates, a, x, e)] affords(s, si, ti, cumulative, a, x) For an example, suppose that the following pattern of interaction holds in situation (saladlab-a, 8): utilises(learner, ingredient(lettuce), wash-ingredient, salad-lab-a, 8) In addition, suppose that the following affordances for patterns of interaction hold in situation type salad-lab-b: affords(salad-lab-b, share, salad-lab-a, 8, ingredient(lettuce)) affords(salad-lab-b, utilises, learner, ingredient(lettuce), add-ingredient-to-salad) Therefore, according to the definition above, we say that the situation type salad-lab-b affords to the learner the development of a course of interaction from situation (salad-lab-a, 8) that is cumulative for the learner with respect to the ingredient lettuce, which means that the following affordance holds: affords(salad-lab-b, salad-lab-a, 8, cumulative, learner, ingredient(lettuce)) Affordance for Self-Regulatedness Self-regulatedness of courses of interaction develops from particular ways in which a learner's action performed in one situation is in some way evaluated by the learner in another situation of the course of interaction, and this evaluation is taken into consideration to guide the next learner's actions or change the effects of previous actions. To afford to a learner the development of courses of interaction that are self-regulated, a situation type must contain particular content entities and afford to the learner the development of particular patterns of interaction that will lead to particular kinds of self-regulatedness being developed. Below is one definition. Definition (Affords self-regulatedness with respect to a content entity from access to context after the event) A situation type s affords to a learner a the development of a course of interaction from situation (si, ti) that is self-regulated for the learner with respect to a content entity x, which is generated through an event e, from access to an evaluation context c, if the learner generates the entity x through the event e in (si, ti), and the entity c is defined in situation type s as an evaluation context for e, and the situation type s affords to the learner accessing the context c through event ec. generates(a, x, e, si, ti) define(context(e, c), s) affords(s, accesses, a, c, ec) affords(s, si, ti, self-regulated, a, x, e, c) For an example, suppose that the following pattern of interaction holds in situation (saladlab-a, 4). generates(learner, part(salad, ingredient(lettuce)), add-ingredient-to-salad, salad-lab-a, 4) Furthermore, suppose that the following context is defined in situation type salad-lab-b for the event type add-ingredient-to-salad, and the following affordance holds: define(context(add-ingredient-to-salad, book("Well balanced salads")), salad-lab-b) affords(salad-lab-b, accesses, learner, book("Well balanced salads"), view-book) 365 Akhras and Self Then, according to the definition above, we say that the situation type salad-lab-b affords to the learner the development of a course of interaction from situation (salad-lab-a, 4) that is selfregulated for the learner with respect to adding an ingredient lettuce to a salad, in the context provided by the book "Well balanced salads", which is the same as saying that the following affordance holds: affords(salad-lab-b, salad-lab-a, 4, self-regulated, learner, part(salad, ingredient(lettuce)), add-ingredient-to-salad, book("Well balanced salads")) INCENSE Based on the theoretical models described in the previous sections, we have implemented INCENSE – an INtelligent Constructivist ENvironment for Software Engineering learning. INCENSE is capable of analysing a time-extended process of interaction between a learner and a set of software engineering situations provided by the environment, in terms of its cumulativeness, constructiveness, self-regulatedness, and reflectiveness. It can then adapt the space of interaction provided by the environment in order to make available to the learner the types of situations that afford the development of further courses of interaction that lead to the desired properties holding. The Domain of INCENSE The domain of INCENSE includes three main phases of software engineering activities: software project planning, software requirements specification, and software design. The general setting is a software engineering laboratory in which two main needs shall arise: Modelling a software engineering process Which involves the learner constructing a model of a particular process of software engineering, such as the process of software project planning, so that the model can be applied when there is a need for it in a project of software development. Applying a model of a software engineering process in a project of software development Which involves the learner using a model of a particular process of software engineering, such as a model of the project planning process, as a basis for developing the activities related to this particular process in a project of software development. The situations of INCENSE correspond to particular cases in which modelling a software engineering process, or applying a software engineering process model, or both, are needed. In situations that involve modelling a software engineering process, the model created by the learner is defined in terms of the following concepts: Processes that are part of the model (e.g. specify requirements) Materials used in the processes (e.g. description of the project scope) Results of the processes (e.g. data flow diagram) Contents of materials or results (e.g. data) Sequences of processes (e.g. specify requirements > check consistency) Figure 1 shows the setting of a learning situation of INCENSE for modelling the process of software project planning. 366 System Intelligence in Constructivist Learning Figure 1. The setting of a learning situation of INCENSE In situations that involve applying a model of a software engineering process, such as in planning a software development project, the application is defined in terms of the model being applied. Therefore, the learner's actions, instead of focusing on creating a model, would be actions to: define the scope of the particular project define the work breakdown structure for the particular project define the activity graph for the particular project perform the critical path analysis for the particular project etc. Therefore, an essential feature of INCENSE that overcomes a limitation of SAMPLE in terms of constructive activity is the possibility of constructing a notion in a modelling situation (e.g. the fact that project planning involves defining an activity graph), and then using this notion constructed, which was not given, to construct another notion (e.g. the particular activity graph that is part of planning a particular project). This would correspond, in the salad design situations of SAMPLE, to constructing the notion that a salad is made of ingredients, which is given in SAMPLE, and then constructing the notion that a particular ingredient is part of a particular salad. Learning Situations of INCENSE Software engineering situations of INCENSE have content and dynamics. The content includes sources of information that can be consulted by the learner while modelling software 367 Akhras and Self engineering processes or applying models of software engineering processes. These sources of information also include the details of the model being created in the situation or the application being developed. Physically, these sources of information are archives that are presented in the situations in graphical form and can be opened using the mouse. As illustrated in Figure 1, each situation has two sets of archives: the situation archives, which contain information about the situation types and, in this case, about software project planning; and the learner archives, which contain the information created by the learner during her or his interaction with the situation. The dynamics of INCENSE situations include a set of interaction events that can be activated by the learner using the mouse and which activate procedures that correspond to the several kinds of actions necessary to create models of software engineering processes, as shown in Figure 1 (e.g. create-process, create-material, etc.), or to apply models of software engineering processes. For example, to include the process "specify requirements" with result "data flow diagram" in her model, the learner activates the procedure create-process, which allows to add a process to the model, which is selected from a list of process-concepts, and then activates the procedure create-result, which allows to define a result for a process (in this case, the process created earlier), selecting from a list of information-concepts. These characteristics of the content and dynamics of INCENSE situations correspond to the external view of INCENSE situations, as these characteristics are physically part of the representation of the situation setting that appears in the screen and with which the learner physically interacts. The formal, internal representation of these and other characteristics of the content and dynamics of INCENSE situations is presented below, for the situation type modelrs2, which is a situation type for modelling the process of software requirements specification. The content of this situation type includes three kinds of objects: the set of archives that are the sources of information for modelling the particular software engineering process, and two sets of software engineering concepts: process concepts and information concepts. Process concepts are used to create processes in the model. Information concepts are used to create materials, results or contents in the model. Some of the objects of these kinds defined in modelrs2 are: archive('Requirements analysis') archive('Interaction trace') process-concept('specify requirements') process-concept('check specification completeness') information-concept('project scope') information-concept('data flow diagram') The dynamics of the situation type includes the definition of the events that support modelling, with their preconditions, effects and associated contexts. To create processes in the model being constructed the following event is defined: event(learner, create-process) Its precondition is the existence of a concept that the learner selects from the set of processconcepts defined above, during the execution of the event: pre(create-process, concept(X), 1) Its effects are: the process created, the information that the process is created being introduced in the lists of process-material and process-results that are shown in the corresponding archives, and the update of the trace of modelling that is shown in the interaction trace archive. effect(create-process, process(X), 1) effect(create-process, part(list-process-material(S, T1), in(process(X), S, T1)), 1) effect(create-process, part(list-process-result(S, T1), in(process(X), S, T1)), 1) effect(create-process, part(trace-modelling(S, T1), occurs(create-process, S, T)), 1) 368 System Intelligence in Constructivist Learning Finally, some of the contexts defined for this event type are: context(create-process, archive('Requirements modelling situation')) context(create-process, archive('Requirements development')) Other situation types that are represented in INCENSE include: model-pp1, which is a situation type for modelling the process of software project definition; model-rs1, which is a situation type for modelling the process of software requirements definition; and model-d1, which is a situation type for modelling the process of data design. According to our approach, the engaging of learners in situation types in which they wish to interact is made from the space of interaction, which contains a selection of situation types that are likely to be beneficial to the learner. As we will see later, the analyses of interaction, process, and affordances, all contribute to the creation of the space of interaction. Interacting in INCENSE's Situations Interaction in situations develops by the occurrence of events. After entering a situation type, the occurrence of the events defined in this situation type give rise to situations that characterise a succession of contexts in which the interaction develops. In the example that we have run, after entering the situation type model-rs2 a sequence of learning events was developed involving the content and dynamics of this situation type. Formally, the events that occurred and the changes in the content of the situations developed by the occurrence of these events are described in terms of the entities of situation development defined earlier. For example, the occurrence of an event in situation (model-rs2, 1) in which a learner creates a process "specify requirements" is encoded as below. occurs(create-process, learner, model-rs2, 1) in(pre(create-process, process-concept(specify requirements), 1), model-rs2,1) As a consequence, the effect of the event holds in situation (model-rs2, 2) as well as the process created. This is encoded as below. in(effect(create-process, process(specify requirements), 1), model-rs2, 2) in(process(specify requirements), model-rs2, 2) Table 3. Representation of occurrences of learning events _____________________________________________________________________ time 1 occurs create_process time 2 in process(specify requirements) in part(trace_modelling(model_rs2, 2), occurs(create_process, model_rs2, 1)) occurs access_archive time 3 occurs create_material time 4 in relation(material(concept(project scope), process(specify requirements))) in information(project scope) in part(trace_modelling(model_rs2, 4), occurs(create_material, model_rs2, 3)) occurs access_interaction_trace time 5 occurs create_result time 6 in relation(result(concept(data flow diagram), process(specify requirements))) in information(data flow diagram) in part(trace_modelling(model_rs2, 6), occurs(create_result, model_rs2, 5)) 369 Akhras and Self In Table 3 we present, in a simplified form, the main entities of situation development generated in the whole interaction (six steps). The fragment of the software engineering process model constructed in these few events is shown in Figure 2. In practical terms, what happened in these six steps of interaction is that the learner started to create her or his model of the process of software requirements specification, whose partial product is illustrated in Figure 2. This model fragment represents that the activity of specifying the software requirements uses the description of the project scope as material and produces a data flow diagram as result. In the course of this interaction, the learner has also accessed the archive that contains information about how software requirements are analysed. Although this archive contains information that helps the learner to build her or his own model, the meaning of entities such as "project scope", "specify requirements", and "data flow diagram" has to be constructed by the learner as she or he applies her or his model in other situations. project scope specify requirement s dat a f low diagram Figure 2. Fragment of a software engineering process model being constructed Analysing Interactions In INCENSE, after the learner leaves the situation type in which she or he has been interacting, the system analyses the interactions to determine the patterns of interaction that have been developed, which correspond to the system's interpretation of the learner's interactions in the situations. The procedures for doing this in INCENSE, implemented in prolog, follow quite straightforward from the formal definitions of the patterns of interaction. An extract is presented below. After the learner A has interacted in situation S from Tm to Tn For each T from Tm to Tn For all X, E such that in(pre(E, X, P),S, T) And for the learner A such that occurs(E, A, S, T) The system asserts that utilises(A, X, E, S, T) holds In addition, if in(X, S, T) and for all Ti<T and Tj<T-1 such that neither utilises(A, X, E, S, Ti) nor generates(A, X, E, S, Tj) The system asserts that new(A, X, S, T) holds Otherwise, it asserts that old(A, X, S, T) holds Similar procedures are used to obtain the other patterns. Figure 3 shows the system's interpretation of the interactions developed in our example concerning the patterns of learner's actions in situations (utilises, generates, and accesses) and the patterns of cognitive states in situations (new and old). In the Figure, software engineering concepts are represented by wiggly bubbles, processes by round bubbles, materials or results by rectangles, and the arrows indicate input and output of learner's events. 370 System Intelligence in Constructivist Learning t im e 1 t im e 2 specif y r equi remen t s specif y re quir ement s ut il ises ge nerat e s new Req uir ement s analysis n ew accesses t ime 4 t im e 3 Trace p ro ject scop e specif y r equi remen t s ut il ises new ut ilises > -- > -- > -- - p ro ject scop e ne w o ld accesse s gen erat e s t im e 6 t im e 5 d at a flo w d iagr am ut ili ses new ut ilises spe ci fy r equ ireme nt s d at a flo w d iagr am n ew old ge nerat e s Figure 3. Patterns of interaction in the situation for modelling the process of requirements specification Analysing the Time-Extended Process of Interaction Following the analysis of interaction, INCENSE proceeds with the analysis of courses of interaction to determine the properties of cumulativeness, constructiveness, reflectiveness, and self-regulatedness, which correspond to the system's interpretation of the time-extended process of interaction developed. In reasoning about courses of interaction, INCENSE uses its knowledge about properties of courses of interaction that we have formally described earlier to obtain the properties that hold in a process of interaction. Table 4 shows some of the properties developed from the interactions of our example. 371 Akhras and Self Table 4. Properties of courses of interaction in the situation for modelling the process of requirements specification _____________________________________________________________________ time 1 time 2 self-regulated with respect to process(specify requirements) and context archive(Requirements analysis) time 2 time 3 cumulative with respect to process(specify requirements) time 3 time 4 constructive with respect to relation(material(concept(project scope), process(specify requirements))) constructive with respect to information(project scope) reflective with respect to trace_modelling time 3 time 6 constructive with respect to relation(result(concept(data flow diagram), process(specify requirements))) constructive with respect to information(data flow diagram) _____________________________________________________________________ Determining the Affordances of Situation Types After determining the patterns of interaction and the process-related properties of a sequence of learning events developed in the situation that the learner has just left, INCENSE analyses the situation types that are part of the environment so that the affordances of these situations to the learner can be determined. As we have discussed earlier, this is done taking into consideration the interactions developed in the situations up to the current time and the possibilities of the new situation types for the development of further interactions that lead to properties of cumulativeness, constructiveness, self-regulatedness, and reflectiveness holding in the following learning interactions. In Table 5 we show some of these affordances for situation type modelrs2. (Note that if an entity contains a variable, such as in the case of process(X), this variable is bound to the name "Var" which then means that the affordance refers to an unbounded entity.) Table 5. Affordances for properties of courses of interaction in situation type model-rs2 _____________________________________________________________________ time 1 time n cumulative with respect to concept(specify requirements) time 2 time n cumulative with respect to process(specify requirements) constructive with respect to relation(material(concept(Var), process(Var))) time 3 time n cumulative with respect to concept(project scope) constructive with respect to relation(material(concept(Var), process(Var))) constructive with respect to relation(before(process(Var), process(Var))) ... time 5 time n constructive with respect to information(Var) ... time 6 time n self-regulated with respect to process(Var) and context archive(Requirements development) reflective with respect to trace_modelling 372 System Intelligence in Constructivist Learning In practical terms, what the analysis of affordances presented above tells the system is that if, for example, model-rs2 is the next situation to be selected by the learner in the space of interaction (in case it is there to be selected), then there are possibilities for cumulativeness with respect to the notions of "project scope", "specify requirements", etc., and for constructiveness, self-regulatedness, and reflectiveness with respect to several other entities. Creating a Space of Interaction As a result of the affordances determined above, a set of situation types qualifies as candidates for the next space of interaction. In our example, we can see that all the four situation types afford some of the desired properties of courses of interaction to be developed in the following process. To sort this set of situation types in terms of their potential benefit for the learner we have defined the following policies: Cumulativeness policy: consider situation types that afford more first-time cumulations. Constructiveness policy: consider situation types that afford more constructions. Self-regulatedness policy: consider situation types that afford more first-time selfregulations. Reflectiveness policy: consider situation types that afford reflections about entities that have had less time spent on them by reflection in previous situations. The results of the analysis of affordances in relation to the policies, for our example, are summarised in Table 6. As the affordances for reflectiveness and self-regulatedness did not change from one situation type to another due to the situation types of the example being all for modelling and having similar mechanisms for self-regulation (access to similar archives), the selection of situation types for the space of interaction in our example was based only on the results of the analysis of affordances in relation to the policies concerning cumulativeness and constructiveness, whose summary is shown in Table 6, in terms of the number of entities that the situation types afford a first time cumulation and the number of entities that the situation types afford construction. Table 6. Summary of affordances situation type cumulativeness constructiveness model-pp2 model-rs1 model-rs2 model-d1 2 3 7 1 11 11 32 5 Therefore, following from the selection of the situation type model-rs2 in the first space of interaction, which contained the situation types model-pp2 and model-rs2, and according to the sequence of learning events developed in the situations derived from the situation type modelrs2, the next two situation types that are more likely to enable, for the learner, the development of courses of interaction that exhibit the desired properties, are the situation types model-rs2 and model-rs1. In practical terms this means that the objects and the possibilities for action that the learner will encounter in the contexts provided by these two types of situations allow the development of learning interactions that make connections with aspects of interactions developed in previous situations and, therefore, help to ensure a continuing learning experience that is meaningful for the learner, as a process. INCENSE Used by Real Students In order to observe some students using INCENSE we have carried out a brief study in which two students used INCENSE, each one interacting individually for about thirty minutes with the 373 Akhras and Self system, producing courses of interaction that extended over fifty events, in which they had to model a process of software engineering project planning. The main goal of this study was to observe the development of properties of cumulativeness, constructiveness, self-regulatedness and reflectiveness in more extended sequences of learning events developed by real students. The results are summarised in the table below. number of different entities cumulated number of different entities constructed number of different entities self-regulated number of different entities reflected upon student A 25 32 1 0 student B 8 22 9 0 Neither student developed reflective courses of interaction, which in INCENSE means that the interaction traces were not accessed. Concerning the properties developed, the numbers show that student A cumulated and constructed more than student B but spent less time selfregulating, which in INCENSE corresponds to accessing archives that contain information about software engineering concepts. Following an analysis of affordances, the system might tend to offer to student A situations that allow fewer new constructions and more self-regulation of the constructions made before, while to student B the system might offer situations that allow more new constructions. CONCLUSION In this paper we have argued that the change in perspective provided by constructivist views of learning, in regard to the nature of knowledge, the way learners learn, and the way learning can be promoted, rather than pointing to a move away from the idea of system intelligence in the computational support to constructivist learning, points to a new kind of system intelligence that is better attuned to constructivist views, which stress the importance of the context of learning, the fact that learning involves active interaction, and the process rather than the product of learning. Focusing on these issues, we have developed a set of formal mechanisms that support the kinds of knowledge representation, reasoning, and decision making that are necessary in order that a constructivist learning environment can develop an evaluation of learning with a focus on the process rather than on the product of learning and can then use this evaluation in order to adapt the environment to the learner needs, aiming for the development of further processes of learning that possess certain desired properties rather than the acquisition of a target knowledge. The whole approach was implemented in INCENSE, which is an intelligent learning environment for software engineering learning that follows a constructivist perspective. INCENSE is able to evaluate learning processes in terms of the four main properties that have been put forward by constructivist theorists as conducive to learning: cumulativeness, constructiveness, self-regulatedness, and reflectiveness. It is also able to adapt the learning environment in a way that facilitates the occurrence of these properties in following courses of interaction. Acknowledgements Thanks are due to the National Council of Scientific and Technological Development (CNPq), Brazil, for sponsoring the research reported in this paper. References Agre, P. E. (1995). Computational research on interaction and agency. Artificial Intelligence, 72, pp. 1-52. Akhras, F. N. (1997). Reasoning about process in intelligent learning environments. PhD. Thesis, Computer Based Learning Unit, University of Leeds. 374 System Intelligence in Constructivist Learning Akman, V. and Surav, M. (1996). Steps toward formalizing context. Artificial Intelligence Magazine, Fall, pp. 55-72. Altman, I. and Rogoff, B. (1987). World views in psychology: trait, interactional, organismic, and transactional perspectives. In: Stokols, D. and Altman, I. (eds.), Handbook of Environmental Psychology, John Wiley, pp. 7-40. Barwise, J. and Perry, J. (1983). Situations and Attitudes. Cambridge, MA: MIT Press. Bednar, A. K., Cunningham, D., Duffy, T. M. and Perry, J. D. (1992). Theory into practice: how do we link? In: Duffy, T. M. and Jonassen, D. H. (eds.), Constructivism and the Technology of Instruction: a Conversation, Hillsdale, NJ: Lawrence Erlbaum, pp. 17-34. Boud, D., Keogh, R. and Walker, D. (1985). Promoting reflection in learning: a model. In: Boud, D., Keogh, R. and Walker, D. (eds.), Reflection: Turning Experience into Learning, London: Kogan Page, pp. 18-40. Brown, J. S., Collins, A. and Duguid, P. (1989). Situated cognition and the culture of learning. Educational Researcher, 18(1), pp. 32-42. Davis, E. (1990). Representations of Commonsense Knowledge. San Mateo, CA: Morgan Kaufmann. De Corte, E. (1995). Learning theory and instructional science. In: Reimann, P. and Spada, H. (eds.), Learning in Humans and Machines, Oxford, UK: Elsevier Science, pp. 97-108. Derry, S. J. and Lajoie, S. P. (1993). A middle camp for (un)intelligent instructional computing: an introduction. In: Lajoie, S. P. and Derry, S. J. (eds.), Computers as Cognitive Tools, Hillsdale, NJ: Lawrence Erlbaum, pp. 1-14. Devlin, K. (1991). Logic and Information. Cambridge University Press. Dewey, J. (1933). How We Think: A Restatement of the Relation of Reflective Thinking to the Educative Process. Boston: Heath. Ertmer, P. A. and Newby, T. J. (1996). The expert learner: strategic, self-regulated, and reflective. Instructional Science, 24(1), pp. 1-24. Gibson, J. J. (1977). The theory of affordances. In: Shaw, R. and Bransford, J. (eds.), Perceiving, Acting, and Knowing, Hillsdale, NJ: John Wiley, pp. 67-82. Greeno, J. G. (1989). A perspective on thinking. American Psychologist, 44(2), pp. 134-141. Greeno, J. G. (1994). Gibson's affordances. Psychological Review, 101(2), pp. 336-342. Greeno, J. G. (1997). On claims that answer the wrong questions. Educational Researcher, 26(1), pp. 5-17. Greeno, J. G., Moore, J. L. and Smith, D. R. (1993). Transfer of situated learning. In: Detterman, D. K. and Sternberg, R. J. (eds.), Transfer on Trial: Intelligence, Cognition, and Instruction, Norwood, NJ: Ablex, pp. 99-167. Hayes, P. J. (1985). The second naive physics manifesto. In: Hobbs, J. R. and Moore, R. C. (eds.), Formal Theories of the Commonsense World, Norwood, NJ: Ablex, pp. 1-36. Kommers, P. A. M., Lenting, B. F., and van der Veer, C. G. (1996). Telematic learning support and its potential for collaborative learning with new paradigms and conceptual mapping tools. In: Proceedings of the European Conference on Artificial Intelligence in Education (EuroAIED), Lisbon, Portugal, pp. 408-414. McCarthy, J. and Hayes, P. J. (1969). Some philosophical problems from the standpoint of artificial intelligence. In: Mitchie, D. (ed.), Machine Intelligence 4, Edinburgh: Edinburgh University Press, pp. 463-502. Ohsawa, I. and Nakashima, H. (1991). Actions in situation theory. Technical Report TR-91-13, Electrotechnical Laboratory, Tsukuba, Japan. Piaget, J. and Garcia, R. (1991). Toward a Logic of Meanings. Hillsdale, NJ: Lawrence Erlbaum. Reiter, R. (1991). The frame problem in the situation calculus: A simple solution (sometimes) and a completeness result for goal regression. In: Lifshitz, V. (ed.), Artificial Intelligence and Mathematical Theory of Computation: Papers in Honor of John McCarthy, Academic Press, pp. 359-380. Resnick, L. B. (1987). Learning in school and out. Educational Researcher, 16(9), pp. 13-20. Resnick, L. B. (1996). Situated rationalism: the biological and cultural foundations for learning. Prospects, XXVI(1), pp. 37-53. 375 Akhras and Self Saada-Robert, M. and Brun, J. (1996). Transformations of school knowledge: the contributions and extensions of genetic psychology. Prospects, XXVI(1), pp. 25-36. Shuell, T. J. (1986). Cognitive conceptions of learning. Review of Educational Research, 56(4), pp. 411-436. Shuell, T. J. (1992). Designing instructional computing systems for meaningful learning. In: Jones, M. and Winne, P. H. (eds.), Adaptive Learning Environments, Berlin: SpringerVerlag, pp. 19-54. Simons, P. R. J. (1993). Constructive learning: the role of the learner. In: Duffy, T. M., Lowyck, J. and Jonassen, D. H. (eds.), Designing Environments for Constructive Learning, Berlin: Springer-Verlag, pp. 291-313. von Glasersfeld, E. (1995). Radical Constructivism: a way of knowing and learning. London: The Falmer Press. von Glasersfeld, E. (1996). Introduction: aspects of constructivism. In: Fosnot, C. T. (ed.), Constructivism: Theory, Perspectives, and Practice, New York: Teachers College Press, pp. 3-7. Vosniadou, S. (1996). Towards a revised cognitive psychology for new advances in learning and instruction. Learning and Instruction, 6(2), pp. 95-109. Winne, P. H. and Hadwin, A. F. (1998). Studying as self-regulated learning. In: Hacker, D. J., Dunlosky, J. and Graesser, A. C. (eds.), Metacognition in Educational Theory and Practice, Hillsdale, NJ: Lawrence Erlbaum, pp. 279-306. 376