UbiCom Book Slides Chapter 8 Intelligent Systems (Part A: Basics) Stefan Poslad http://www.eecs.qmul.ac.uk/people/stefan/ubicom Ubiquitous computing: smart devices, environments and interaction 1 Chapter 8: Overview Chapter 8 focuses on: • Internal system properties: intelligence • External interaction with any of three types of environment – Focussing more on ICT and physical environment – These environments tend to be passive Ubiquitous computing: smart devices, environments and interaction 2 Five main properties for UbiCom Ubiquitous computing: smart devices, environments and interaction 3 Ubiquitous computing: smart devices, environments and interaction 4 Chapter 8: Overview The slides for this chapter in the full pact are expanded and split into several parts • Part A: Basics • Part B: R-IS, EM-IS • Part C: G-IS, U-IS & H-IS Models • Part D: KB IS Models • Part E: KB Acquisition • Part F: KB Representation: Rule-Based, BB • Part G: KB Representation: Semantic • Part H: Classic Logic KB Models • Part I: Soft Computing KB Models • Part I: Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 5 Ubiquitous computing: smart devices, environments and interaction 6 Related Chapter Links • There are two AI chapters that are interlinked • This chapter. 8, describes the design of single Intelligent System or IS – These may be simple: use a single models of intelligence – These may be hybrid: use multiple heterogeneous intelligence models • Chapter 9, describes the interaction between multiple ISs Ubiquitous computing: smart devices, environments and interaction 7 Related Chapter Links • Many AI researchers see autonomy as a sub-type of intelligence. – Sometimes their notion of autonomy is not well-defined • In this text we separately autonomy and intelligence as main concepts – both as main types of property for UbiCom • A very rich model for autonomy is given in Chapter 10 – Relates the UbiCom properties of intelligence & autonomy Ubiquitous computing: smart devices, environments and interaction 8 Related Chapter Links • Reflexive Intelligent system models are related to: – – – – Event-based system models (Chapter 3) Sensor systems for the physical Environment (Chapter 6) Controller systems for the physical Environment (Chapter 6) Context-aware system design (Chapter 7) • Hybrid Goal-based environment model: – Based upon used of EDA, BB, pipe-filter interaction design (Chapter 3) – Used in context-aware system design (Chapter 7) – Used in autonomic computing designs (Chapter 10) • Monitoring & analysis of self rather than one’s environment – Self-awareness & Reflection (Chapter 10) Ubiquitous computing: smart devices, environments and interaction 9 Overview • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics Logic reasoning Models for IS Soft Computing IS Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 10 Intelligent systems (IS): Introduction • Systems that use AI algorithms. • Also called – Machine intelligence or computational intelligence – Agent-based systems • AI effect ??? Ubiquitous computing: smart devices, environments and interaction 11 IS Introduction • IS Models often use analogies of human problem solving • IS models can be based upon physical organisations • But Machines may not problem solve like humans Ubiquitous computing: smart devices, environments and interaction 12 Human Intelligence vs. Machine Intelligence • Compare & Contrast • What complex tasks are humans good at and machines less so and vice versa? Ubiquitous computing: smart devices, environments and interaction 13 Types of Intelligence Can we define intelligence as any single concept in any single definition? Or is intelligence a multi-dimensional concept Ubiquitous computing: smart devices, environments and interaction 14 Ubiquitous computing: smart devices, environments and interaction 15 Intelligent Systems (IS) IS vs. distributed system? Ubiquitous computing: smart devices, environments and interaction 16 UbiCom Systems that use an implicit notion of Intelligence Usage of term AI in general computing varies • HCI • Sensor / Context-aware systems • Control system & Robots • Intelligent networks • Network, e.g., SNMP, Agents • AmI • Smart devices Ubiquitous computing: smart devices, environments and interaction 17 Types of IS Model IS models can be classified in terms of: • Type of Model / Architecture • How a model is used to solve some problem • What is being modelled • What types of environment, a system is situated in, and can operate in • How are the models in an IS acquired? Ubiquitous computing: smart devices, environments and interaction 18 Types of IS Model Representations IS models can be classified in terms of: Basic types of IS Representation • Process-driven system models: • Data-driven KB-IS models: • Logic –based KB-IS models: • Soft Computing models: Ubiquitous computing: smart devices, environments and interaction 19 Unilateral System Environment Interaction Models • Generally when ubiquitous system applications are designed, a unilateral model of the environment is used. Explain this? Ubiquitous computing: smart devices, environments and interaction 20 Unilateral versus Bilateral System Environment Models • Generally when ubiquitous system applications are designed, a unilateral model of the environment is used, – system models its environment, not vice-versa • However, as we move to smarter environments – Environments can be designed to contain a model of application systems which are situated in them or pass through them – Environment can be considered itself as an active, smart system Ubiquitous computing: smart devices, environments and interaction 21 Unilateral versus Bilateral System Environment Models Ubiquitous computing: smart devices, 22 environments and interaction Bilateral System Environment Models • A system that models an active environment • And in turn, the active environment has a model of the systems which use it • Designers of systems may to take into the account the degree of intelligence of environments Ubiquitous computing: smart devices, environments and interaction 23 IS Environment Types • It is a challenge for any system to act in open system environments. • Russell and Norvig (2003) have categorised open system environments along several dimensions . • Simplest types of system environments are those that are fully-observable, episodic and static. • More complex designs for intelligent systems are needed to think and act in environments that: are uncertain and nondeterministic etc. Ubiquitous computing: smart devices, environments and interaction 24 IS System Environment Types Environment Description of Environment Fully observable Deterministic Episodic Static Discrete Passive Antonym UbiCom’s sensors give it access to complete state of Partially environment at each point in time. observable Next state of environment determined by current state Stochastic and action executed by UbiCom system. UbiCom choice of action depends only on current state Sequential of environment on episode itself. Environment static while UbiCom system selects & Dynamic execute its actions i.e., to adapt to its environment A limited number of distinct, clearly defined states and Continuous actions characterise the environment environment ii not active, in sense of modelling the Active system that is acting in it. Ubiquitous computing: smart devices, environments and interaction 25 IS Environment Types • Open system environment are often stochastic. Why? Ubiquitous computing: smart devices, environments and interaction 26 IS Environment Types • Discuss examples of these System Environment Types: Ubiquitous computing: smart devices, environments and interaction 27 What is Modelled and How the Model is Acquired • What is Modelled? • Different ways to acquire the model? Ubiquitous computing: smart devices, environments and interaction 28 When Models are Acquired: At Design Time Contrast acquiring the model itself versus acquiring the content to populate the model 2 main ways to design how the model a acquired? • Models can be built into system at design time & modified at run-time • Models can be built at design time so that they can modify themselves at run-time Ubiquitous computing: smart devices, environments and interaction 29 When Models are Acquired • Models can be created by a human designer & built into system at design time & upgraded later. • What are benefits? – ?? • What are the limitations? – ?? • Systems/ environment can acquire their models themselves, automatically • Hybrid model acquisition systems Ubiquitous computing: smart devices, environments and interaction 30 Overview • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics Logic reasoning Models for IS Soft Computing IS Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 31 Should a UbiCom System be an IS? • Depends upon the nature of the application. • Depends upon the specific model of AI being used • Specific IS models can be used to build systems which support other UbiCom system properties such as contextawareness, autonomy and iHCI Ubiquitous computing: smart devices, environments and interaction 32 Basic Types of IS Model • • • • • • • (From Russel & Norvig 2002) Reflex Based Environment Model based Goal based, Proactive Utility based Learning Multi-IS, Hybrid IS Ubiquitous computing: smart devices, environments and interaction 33 Basic Types of IS Model Note also they can classify the types of IS model by the knowledge representation they use for their model • Rule-based • Light-weight Ontology • Heavy-weight Ontology • Active user versus active service processes Ubiquitous computing: smart devices, environments and interaction 34 Types of IS Model & Types of Environment They Suit Type of Model IS What a IS’s actions Types of environments IS design is depend upon suited to Reflex Based Current Environment context Env. Model Current and past based, Situated environment context action Goal based, IS’s plans of actions Proactive to achieve a goal Utility based IS’s weighting of different goals and plans. Learning Performance Multi-IS, Hybrid IS Fully-observable, stochastic, episodic, static, physical Partially-observable, deterministic, sequential, dynamic Partially-observable, deterministic, sequential, dynamic human Partially-observable, Semideterministic, sequential, dynamic Partially-observable, Nondeterministic, sequential, dynamic Partially-observable, Nondeterministic, sequential, dynamic Ubiquitous computing: smart devices, environments and interaction 35 Reactive IS Models (R-IS) • Intelligent behaviour arises out of system’s interaction with environment rather than as result of complex internal knowledge representation or reasoning about events. • Action selection is at heart of the intelligent system. – in the simple case is driven by current state of environment. • R-IS is strongly situated in its environment and is highly responsive to changes in the environment. Ubiquitous computing: smart devices, environments and interaction 36 Pure versus Hybrid R-IS Models • R-IS tend to be designed as event-based systems – • Pure reactive type of IS works best when? – • In practice many systems are designed not to be purely reactive – • These represent hybrid reactive systems. Ubiquitous computing: smart devices, environments and interaction 37 (Pure) R-IS Model Ubiquitous computing: smart devices, environments and interaction 38 R-IS Design: Present action Rulebased actions • Preset actions may be directly triggered from sensor input without any conditions • Alternatively events can be filtered by conditions / rules in order to trigger actions Ubiquitous computing: smart devices, environments and interaction 39 R-IS Design: handling multiple concurrent & heterogeneous events 3 possible designs: • Discard events • Event persistence • Concurrent Event Handling Ubiquitous computing: smart devices, environments and interaction 40 R-IS Design: handling multiple concurrent & heterogeneous events Ubiquitous computing: smart devices, environments and interaction 41 UbiCom Systems based upon R-IS • R-IS design is good design for minimum context-aware CA) system Ubiquitous computing: smart devices, environments and interaction 42 Environment Model based IS (EM-IS) • EM-IS ~ KB system: knowledge about world & its actions. • EM-IS models historical behaviour of its environment. • System environment may be partially observable. Why? Ubiquitous computing: smart devices, environments and interaction 43 EM-IS • How to handle partial observability of environment? Ubiquitous computing: smart devices, environments and interaction 44 EM-IS Ubiquitous computing: smart devices, environments and interaction 45 EM-IS • IS’s actions depend upon current environment state, past environment states & on knowing effect of system actions • Similar to a situated action type of system design: – actions can be unplanned and depend strongly on context • Can anticipate multiple future environment states, which may never be realised leading to a theory of multiple possible future environments or world states. • EM-IS not include a model of the internal behaviour, e.g., processes of actions by the system Ubiquitous computing: smart devices, environments and interaction 46 EM-IS • Systems that build such a model of the environment enable their services & applications to optimise & adapt their behaviour to account for behaviours in environment which are not accessible but which are predetermined (as defined in the environment model). • E.g.,, in the adaptive transport scheduling scenario, – In pure R-IS, vehicle will not stop when no passengers are at pickup point (providing no passengers on the vehicle wish to leave). – EM-IS: ….. Ubiquitous computing: smart devices, environments and interaction 47 Overview • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrid IS Models KB-IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics Logic reasoning Models for IS Soft Computing IS Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 48 Goal-based IS (G-IS) • Also referred to as planning-based IS, defines an internal plan or sequence of actions to achieve a future system goal • Unlike EM-IS, action selection for a G-IS depends on which next system action brings system towards future goal state • G-IS tends to dissociate control of actions from environment situation or context of action (unlike EM-IS) • G-IS vs. R-IS, events which trigger system actions as external events • In G-IS, internal events, e.g., a scheduled system task that becomes delayed, can also trigger system actions. • Main benefit of G-IS: users can delegate tasks at a much higher level of abstraction, focussing on what needs to be achieved rather than on details of how this is achieved Ubiquitous computing: smart devices, environments and interaction 49 G-IS Ubiquitous computing: smart devices, environments and interaction 50 Utility-based IS (U-IS) • Utility refers to quantifiable measure of performance or worth or usefulness of specific goal in set of possible goals. • Can also refer to a specific chain of actions amongst set of possible chains of actions. • Utility function maps (goal) state or a chain of states to value which represents its performance or worth. • U-IS design is useful when: – several conflicting goals exist – multiple goals are possible but only one of them is practical or achievable. Ubiquitous computing: smart devices, environments and interaction 51 U-IS: Adaptive transport scheduling scenario • 2 conflicting goals to recover from disruptions to designated schedule are: – maximise revenue by maximising pickup load maintaining a quality of service – minimising deviations from a designated schedule. • The greater load picked up, the later an already late vehicle remains or gets worse. • A utility function weighs revenue generation & maintaining punctuality – E.g., ?? Ubiquitous computing: smart devices, environments and interaction 52 Overview • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB-IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics Logic reasoning Models for IS Soft Computing IS Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 53 Hybrid IS (H-IS) • H-IS models are more complex and aim to combine the benefits of the individual IS models. • 2 basic designs: – Horizontal concurrent layers – vertical (sequential) layers. • Layers consist of single or multiple IS components with a clearly defined interface for input and output. Ubiquitous computing: smart devices, environments and interaction 54 H-IS: Vertical vs. Horizontal Model Designs Ubiquitous computing: smart devices, environments and interaction 55 Hybrid IS (H-IS): Horizontal Layers • Horizontal homogeneous layered model (Section 8.3.2) allows multiple events to be handled in parallel. • This type of model just needs to be generalised to allow heterogeneous models to be layered. • H-IS models can be ‘layered’ in a single IS • H-IS can handle reactive events in lower reactive layer • Then handle events which require use of environment model & reasoning in a higher layer. Ubiquitous computing: smart devices, environments and interaction 56 H-IS: Horizontal Layered Model • Design challenge with this type of model is that multiple output action events can occur for the same input event. – because each layer independently outputs its own action. • Problems ? • Solutions? Ubiquitous computing: smart devices, environments and interaction 57 H-IS: Vertical Layered Model • Simplest chaining is a type of single-pass vertical model designs in which control flows through each layer in order to generate the action in the last layer. • Other types of vertical model design? – could use multiple passes or flows for control, information and action generation. • Challenges? – Does not allow concurrent event processing tasks to occur and can form a processing bottleneck. – If any component fails the whole chain could fail without design support to prevent this. • Solutions? Ubiquitous computing: smart devices, environments and interaction 58 Embedded Control System & Robots based upon R-IS , U-IS • Simple control systems could be based upon R-IS & U-IS • Discuss … Ubiquitous computing: smart devices, environments and interaction 59 Hybrid L-IS, G-IS Vs. Autonomous Systems • Autonomous systems are similar to IS in the sense that these can be goal-directed and policy constrained. • Discuss … Ubiquitous computing: smart devices, environments and interaction 60 H-IS Applications Can Operate in Multiple Environments • UbiCom systems are sometime needed to operate in multiple environments that are chained • Process events in model-based human environment, then in goal-based model, then in human reactive model and finally in a physical world reactive model. • e.g., in adaptive transport scheduling scenario: – transport system needs to be aware of human environment. Why? – Transport system needs to be aware of phys. Env. Why? – Transport system may be driven by goals & utility functions Why?. Ubiquitous computing: smart devices, environments and interaction 61 Hybrid IS: Vertical Layered Model • Example of use of hybrid IS Vertical Layered Model for UbiCom that models reactive human events, reactive physical events, model of human environment and models of goals. Ubiquitous computing: smart devices, environments and interaction 62 Hybrid IS: Horizontal Layered Model Ubiquitous computing: smart devices, environments and interaction 63 Hybrid R-IS, EM-IS and U-IS Application: Adaptive Transport Scheduling scenario • IS represents transport, e.g., bus, service • Use utility functions to weight the importance of different independent factors that affect a goal, – e.g., …. • Prediction of a bus’s arrival at scheduled bus-stops is based upon a model of how a bus’s environment, – e.g., …. • IS to support this scenario could be based upon a hybrid RIS, EM-IS and U-IS design. Ubiquitous computing: smart devices, environments and interaction 64 Hybrid R-IS & G-IS Application: Foodstuff Management Scenario • Discuss ... Ubiquitous computing: smart devices, environments and interaction 65 Hybrid R-IS & G-IS Application: Utility Regulation Scenario • Discuss … Ubiquitous computing: smart devices, environments and interaction 66 Overview • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics Logic reasoning Models for IS Soft Computing IS Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 67 Knowledge based (KB) IS Models • A range of IS models in terms of: – representation, – operations – what the KB model is used for (the type of KB model) • Commonly used types of KB system architectures include: – Blackboard systems & EDA systems – Production systems or rule systems – Semantic type KBs such as ontology-based systems (Section 8.4) • Alternative Terms are sometimes used somewhat synonymously by some is this correct? – knowledge-based, – semantic – ontology-based models Ubiquitous computing: smart devices, environments and interaction 68 KB Models: Benefits • • • • • to share a common understanding to enable the reuse of domain knowledge, to make domain assumptions explicit, to separate domain from operational knowledge to analyse domain knowledge Ubiquitous computing: smart devices, environments and interaction 69 KB Characteristics • KB behaves as a surrogate • KB is a set of ontological commitments • KB is often used for intelligent reasoning • KR acts as a machine readable & understandable language • KR acts as a human readable &understandable language Ubiquitous computing: smart devices, environments and interaction 70 What is Knowledge? • • • • Know-what? Know-how? Know-why? Experiences? Ubiquitous computing: smart devices, environments and interaction 71 KB IS Management Life-cycle • KB Manual Creation vs. Automatic Acquisition • KB Deployment & Maintenance – Validating – Updating in a consistent way • KB Management & services: store, retrieve, share KB models Ubiquitous computing: smart devices, environments and interaction 72 Why ISs Use Knowledge Models 1. EM-IS: Knowledge and reasoning play a crucial role in dealing partially-observed environments 2. EM-IS: Knowledge and reasoning play a crucial role in dealing with sequential environments: 3. Multi-ISs: enhances interoperability 4. Multi-ISs: enriches information and task sharing 5. Multi-ISs: enable reuse of domain knowledge Ubiquitous computing: smart devices, environments and interaction 73 Overview • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics Logic reasoning Models for IS Soft Computing IS Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 74 What is Modelled in each IS? • • • • R-IS: EDA rules EM-IS: EM G-IS: goals and plans U-IS: weightings or goals or tasks • Seems a useful model for representing personal preferences – When and how are these model defined? – When how are they populated with information? Ubiquitous computing: smart devices, environments and interaction 75 How the models in an IS are acquired • Defined at fixed design time by humans. Data to populate the model can be acquired at run-time: – From humans – From other ISs, knowledge-sharing (Chapter 9) – From ‘non-intelligent ‘data repositories, e.g., databases (Chapter 3) • Defined at design time by humans so that models can improve themselves – i.e., incorporate learning – Any of the models in any type of IS can be learn Ubiquitous computing: smart devices, environments and interaction 76 Knowledge Life-cycle & Management: Creating Knowledge • In order to create a useful & accurate application domainbased knowledge based system need combination of: – an understanding of the problem – a collection of heuristic problem-solving rules from experience • Earliest type of KB system tended to use the knowledge of one or more human domain experts for source of a system's problem solving strategies – Often referred to as an expert system. – E.g., ??? – etc Ubiquitous computing: smart devices, environments and interaction 77 Knowledge Life-cycle & Management: Creating Knowledge • KB consists of concepts and their properties , concept relationships and property relationships • A concept represents an idea of something that could be a real world object or an abstract object such as human behaviour, or feelings etc. • Creating an Ontology for devices, e.g., all device concepts that have: – either a microprocessor, microcontroller or central processor unit, – have memory, have an input interface, have an output interface, have a network interface, are collectively described as ‘devices’. Ubiquitous computing: smart devices, environments and interaction 78 Example Domain Ontology for Devices Link labelled graph: uses multiple types of link, e.g., W3C RDF, RDF-S, OWL etc Ubiquitous computing: smart devices, environments and interaction 79 Design Issues: Creating a Device Ontology • In general, all such concepts do not have any absolute definition; • E.g., , concept of device is defined in terms of its relationship to pre-existing concepts • Hence, concepts are understood through their relationship to other concepts, which have already been understood and remembered. Ubiquitous computing: smart devices, environments and interaction 80 Creating a Domain Ontology: Manual vs. Automatic Process • Process is manual? – e.g., HCI UCD like • Process is automatic? – E.g., uses machine learning, data mining Ubiquitous computing: smart devices, environments and interaction 81 Creating a Domain Ontology: Process • The process of creating an Ontology for a domain consists of defining: – – – – Concept taxonomy A set of relations Constraints Axioms Ubiquitous computing: smart devices, environments and interaction 82 Creating a Domain Ontology Process: Design Issues • There are many different variations of the creation process. • The ease and explicitness at which complex relationships can be modelled depends upon the type of KR: light-weight versus heavy-weight. There are many modelling choices to be made in Ontology design as in any kind of design: • Categorisation choices • Model relationships which have different cardinality • Modelling the direction & symmetry of relationships • Modelling choices about constraints Ubiquitous computing: smart devices, environments and interaction 83 Ontology Refinement & Validation • Often a knowledge model which is created requires a process or refinement in order to improve it • Ontology Model needs to be validated. Why? • How? • Once an Ontology is defined & agreed, can it vary? Ubiquitous computing: smart devices, environments and interaction 84 Knowledge Life-cycle & Management: Maintaining Knowledge Design issues • Can the knowledge model then remains fixed during deployment? • Or do knowledge models in practice need to change • Or multiple knowledge models may exist ? Ubiquitous computing: smart devices, environments and interaction 85 Knowledge Life-cycle & Management: Maintaining Knowledge In practice, knowledge models may vary within a domain Why? Ubiquitous computing: smart devices, environments and interaction 86 Maintaining Knowledge: Handling Heterogeneous Models • How to deal with multiple heterogeneous knowledge models within a domain? • Interoperation between heterogeneous ontology models – Merging or integration – Alignment Ubiquitous computing: smart devices, environments and interaction 87 Learning-based IS (L-IS) • Learning refers to system improving its performance with experience, with respect to some task. • System is said to learn from experience E with respect to some class of actions A and performance measure P, if its performance at the set of actions A, as measured by P, improves with experience E. • Adaptive transport scheduling example: – A = “a logistics vehicle picks up goods on route”, – E = “travelling the route”, P = “deviation of actual time from predicted time”. – Improvement is measure P reducing to zero. Ubiquitous computing: smart devices, environments and interaction 88 L-IS Ubiquitous computing: smart devices, environments and interaction 89 L-IS • Design of the learning element depends on: – what (which model) is learned, – the type of feedback and the model or knowledge representation. • Learning may need model representations that can handle uncertainty • 3 main types of learning or feedback which can be used: – supervised learning, – unsupervised learning – reinforcement learning. Ubiquitous computing: smart devices, environments and interaction 90 Overview • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge Representations: rules, BB, semantics • Logic reasoning Models for IS • Soft Computing IS • Generic IS Operations Ubiquitous computing: smart devices, environments and interaction 91 Knowledge Representations (KR) Commonly used types of Knowledge Representations: • Blackboard systems & EDA systems • Production systems & rule-based systems • Syntactical: RDBMS, XML Web services • Semantic type KBs such as ontology-based systems • Classic-logic based • Soft Computing Which type of knowledge can these express well, express less well? Ubiquitous computing: smart devices, environments and interaction 92 KB Systems: Production or RuleBased • In a production system, knowledge is represented as a set of rules or productions stored in a KB • Rules are typically defined as if IF-fact THEN-fact – IF-fact (also called the condition part or antecedent part) – THEN-fact (also called the consequent part or action part) • Rule-based KB model can be combined with a reactive type IS. Ubiquitous computing: smart devices, environments and interaction 93 KB Systems: Rule-Based Adaptive Transport Scheduling Service Scenario example Ubiquitous computing: smart devices, environments and interaction 94 KB Systems: Rule-Based • Give some examples of rule-based engines fragments for the other 3 scenarios: – Personal memories – Foodstuff management – Utility Regulation Ubiquitous computing: smart devices, environments and interaction 95 KB Systems • Rules can get added to the KB: – Manually, e.g., enterprise policy-based systems – Automatically, e.g., machine learning. • Many rule-based engines have been developed – They enable rule-based systems to incorporated as part of more general distributed systems versus as part of more specialised IS, – e.g., JESS Ubiquitous computing: smart devices, environments and interaction 96 Rule-based KB Systems: Challenges • ??? Ubiquitous computing: smart devices, environments and interaction 97 Rule-based KB Systems • Rule-based KB vs. Semantic KB? Ubiquitous computing: smart devices, environments and interaction 98 KB System: Blackboard (BB) • BB systems act as a shared knowledge type data repository – For use between multiple possibly distributed processes – See Section 3.3.3.7 • Knowledge sources can be: – Independent & distributed – heterogeneous Ubiquitous computing: smart devices, environments and interaction 99 KB System: BB vs. EB KB vs. Event-based system (EB)? Ubiquitous computing: smart devices, environments and interaction 100 Semantic IS: KRs • Ontology based models support a rich semantic conceptualisation & can directly support reasoning – Huge amount of research interest, less mainstream industry use – but this depends on how terms ontology, semantics are defined. • Often term Ontology, expressed informally as a collection of descriptions of the world that helps users define the meaning of their actions on the world, is used synonymously with the term KR • Ontologies have many more formal definitions – e.g., “A formal, explicit specifications of a shared conceptualization” – etc Ubiquitous computing: smart devices, environments and interaction 101 KR: Multiple Semantic Representations: Light-Weight to Heavy-Weight • There exists a range of ontology models and representations depending how concepts and their relationships are defined and organised. • Light-weight • Medium-weight • Heavy-weight Ubiquitous computing: smart devices, environments and interaction 102 KR: Light-Weight Representations • Have simple conceptualisation having parts such as values of terms that may not be machine-readable and machinerelatable to other terms. – E.g., dictionaries • Currently, the most widely used light-weight KRs based upon W3C Web XML • Defines an unnamed hierarchy of concepts & properties • Acts as a basic node labelled graph representation Ubiquitous computing: smart devices, environments and interaction 103 Light-Weight KR: Node-Labelled Graph Node labelled graph: uses one type of inter-node link and intra-node link, e.g., W3C XML Ubiquitous computing: smart devices, environments and interaction 104 KR: Light-Weight Representations • XML is an extensible language designed for exchanging extensible application specific hierarchical data structures. • XML extensions Is used for SOC (Section 3.2.4) • Data structures are certainly machine readable • XML is a difficult data format on which to build automated machine-understandable processing and to support interoperability between autonomous heterogeneous Web services. • Extensions to this are heavier weight and developed as part of the Semantic Web (SW). Ubiquitous computing: smart devices, environments and interaction 105 Medium-Weight KR: Edge Labelled Graph Link labelled graph: uses multiple types of link, e.g., W3C RDF, RDF-S, OWL etc Ubiquitous computing: smart devices, environments and interaction 106 KR: Heavy-Weight Representations • These support more descriptive conceptualisation and more expressive constraints on terms and their interrelationships including logical constraints • Regardless of the properties of the specific Ontology, heavier-weight Ontologies generally include: – ???. Ubiquitous computing: smart devices, environments and interaction 107 KR: Heavy-Weight Representations: Which Types of Logic • Heavy-weight semantic KRs that are also logical KRs • Which type of logic should be combined with the semantic representation to support? – • How to support more flexible reasoning? Ubiquitous computing: smart devices, environments and interaction 108 Semantic Web • Semantic Web (SW) was created: – to evolve the Web from machine-readable to > machineunderstandable – to support richer service interoperability • SW defines a suite of KRs with different weights – RDF – RDFS – OWL Ubiquitous computing: smart devices, environments and interaction 109 Semantic KR: Design Issues • Open World versus Closed World Semantics • Knowledge Life-cycle and Knowledge Management – Creating Knowledge – Knowledge Deployment – Maintaining Knowledge • Design Issues for UbiCom Use Ubiquitous computing: smart devices, environments and interaction 110 KR Design: Open World versus Closed World Semantics • RDBMS type KB models tend to assume closed-world semantics, – if data is not present, it is false (negation as failure). • In contrast semantic type KBs tend to use open-world semantics – regard absence as unknown (negation as unknown). • Is there an Open World versus Closed World Semantics clash? Ubiquitous computing: smart devices, environments and interaction 111 KR Design: Semantics is Variable or Undefined • Semantics for the same concept may be dynamic, depend on context • Understanding of semantics varies across heterogeneous users • Semantics of the data is often defined not to be declarative – ?? • Application semantics are often not explicit and not accessible from the data • Solution – ?? Ubiquitous computing: smart devices, environments and interaction 112 KB IS: Creation Tools • Several interactive knowledge creation tools available • Enable less expert and specialised developers & users to create knowledge models • Then to export the knowledge model out of the tool in a form that can be imported, interpreted or parsed and then invoked via an API by computer applications – E.g., Protégé – E.g., Jena Ubiquitous computing: smart devices, environments and interaction 113 KR Design for UbiCom Use • ICT resources in some devices, particularly, mobile, ASOS and embedded devices are limited by design. • May not be possible to handle semantic information & commands well on the device. • Length of time a computation takes affects the semantics? • Designs for KR for IS operations need to be selected to ?: Ubiquitous computing: smart devices, environments and interaction 114 KR Design for UbiCom Use: Device Only have Access to Partial Models • Devices embedded into a local environment often have partial view rather than a global view of their environment. • Systems today often create information in a far more decentralised manner than they did in the past. Why? – • Solutions? – Ubiquitous computing: smart devices, environments and interaction 115 Lecture Outline • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge representation: rules, BB, semantic Classic Logic IS Models Soft Computing IS Models IS System Operations Ubiquitous computing: smart devices, environments and interaction 116 Classical Logic IS • Propositional and Predicate Logic • Reasoning • Design Issues Ubiquitous computing: smart devices, environments and interaction 117 Classic Logic • Classical logic based upon first-order predicate true or false logic as being at the heart of any IS which needs to support reasoning about the system. • These type of IS are also commonly referred to as (logic) reasoning systems, deliberative IS, and as symbolic AI because these systems involve the manipulation of symbols in the form of logic formulae, although in general symbols could also refer to any mathematical formulae including algebraic formulae. Ubiquitous computing: smart devices, environments and interaction 118 Propositional Logic • Propositional logic : knowledge represented in the form of relations which are either true or false. • Multiple propositions can be combined to form sentences using logic operators that are either true or false • The standard logic operators are: – And, or, not, equals, implies Ubiquitous computing: smart devices, environments and interaction 119 Predicate Logic • Predicates are defined to support more expressive sentences than propositions – allow a property to be related to some object or – a property related to some value. • Sentence "Device A is in hibernate mode", is a predicate • Can be evaluated to a proposition, – e.g., mode (Device A, Hibernate). • Most common form of Predicate logic is called First-Order Predicate Logic or FOPL. Ubiquitous computing: smart devices, environments and interaction 120 Description Logics (DL) • Based Upon combining FOL with a conceptualisation organisation based upon graphs, e.g., RDF-S • Used extensively in Heavy-weight Ontology languages, Semantic Web, e.g., OWL • Benefits? • Consist of ? Ubiquitous computing: smart devices, environments and interaction 121 Reasoning • Reasoning or inferencing, involves logical operations on logical sentences or statements within a (logical) model, bounded by an application domain etc, in order to draw conclusions and to derive other sentences, – e.g., A entails B, A |= B. Inferencing is used to search for entailments. • Sometimes multiple possible worlds or models will be possible. Why not? Ubiquitous computing: smart devices, environments and interaction 122 Reasoning: Model checking • Model checking used to check that entailments of sentences are valid in all possible worlds or models. • Valid sentences are called tautologies. • Sometimes it is just necessary to check if a sentence is true in some specific model, i.e., it is satisfied in that model rather than being able to say it is true in all models, i.e., it is valid in all models. • Model checking can involve changing logical restructuring, changing the syntax of logical sentence whilst keeping the semantics the same, in order to make checking the logical equivalence of two sentences easier. Ubiquitous computing: smart devices, environments and interaction 123 Reasoning: • Resolution type inferencing can result in lengthy computation. • This is particularly an issue in resource constrained devices Solutions? Ubiquitous computing: smart devices, environments and interaction 124 Reasoning: Design Issues For more pervasive use of logic-based IS that supports reasoning etc. • Reasoning needs to be scalable • Reasoning needs to be selectively used • Reasoning when it occurs needs to be computationally efficient. Ubiquitous computing: smart devices, environments and interaction 125 Reasoning: Design Issues Benefits of reasoning using FOL ? Ubiquitous computing: smart devices, environments and interaction 126 KR Systems Based Upon Classic Logic: Challenges • • • • • • Difficulty in expressing exceptions Imprecision Uncertainty High computation maybe needed to establish truth Logical inconsistencies can occur, e..g, in a distributed KB Different sub-types and extensions to classical logic exist Ubiquitous computing: smart devices, environments and interaction 127 Lecture Outline • • • • • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge representation: rules, BB, semantic Classic Logic IS Models Soft Computing IS Models IS System Operations Ubiquitous computing: smart devices, environments and interaction 128 Soft Computing • Many decisions which involve interaction with humans and the physical world are soft – rather than being expressed as either true or false. • Are more qualitative and may involve some imprecision and uncertainty. • Such systems can be designed using soft computing techniques, Ubiquitous computing: smart devices, environments and interaction 129 Probabilistic Networks • Probabilistic network, also called a Belief network or Bayesian Network (BN) • How to model the likelihood of indeterminate events happening or to model the degree of belief in a proposition or predicate and then to reason about them? • Must consider a prior or unconditional probability and conditional or posterior probabilities • Product law expresses a conditional probability in terms of another conditional probability and two unconditional probabilities • This is the basis of probabilistic inferencing. Ubiquitous computing: smart devices, environments and interaction 130 Bayesian Network (BN) • Bayesian Network (BN)can be used to represent any full joint probability distribution. • BN can be used to inference in a context-aware UbiCom scenario in which there are both non-deterministic preconditions and non-deterministic outcome. – E.g., in the adaptive vehicle scheduling scenario, both passengers and buses can indeterminately arrive at pickup points and humans and vehicles can indeterminately wait (Figure 8-11). Ubiquitous computing: smart devices, environments and interaction 131 Bayesian Network application: Adaptive Transport Scheduling Scenario Ubiquitous computing: smart devices, environments and interaction 132 Fuzzy Logic • Fuzzy logic can represent a where model – the outcome of a proposition is deterministic – but is somewhat approximate or imprecise – E.g., the vehicle is travelling very slowly, or slowly, or at a moderate speed, or fast or very fast. • This kind of imprecision can also be used in fuzzy rules. Ubiquitous computing: smart devices, environments and interaction 133 Fuzzy Logic Application: Adaptive Transport Scheduling Scenario • Example fuzzy logic rule could be – If the bus is travelling slowly away from the pickup point – and a passenger is moving quickly towards the pickup point – then slow down the vehicle to stop near the pickup point. • Here, the terms slowly, quickly, and near, act as fuzzy descriptors. Ubiquitous computing: smart devices, environments and interaction 134 Fuzzy Logic Application: Adaptive Transport Scheduling Scenario • Why wouldn’t crisp logic work here? Ubiquitous computing: smart devices, environments and interaction 135 Lecture Outline • • • • • • Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB Management: Creation & Deployment Knowledge Representation: rules, BB, semantic • Classic Logic IS Models • Soft Computing IS Models • IS System Operations Ubiquitous computing: smart devices, environments and interaction 136 IS System Operations • Searching • Planning • Reasoning but is specific to a specific logic representation • Learning: requires architectural support for feedback Ubiquitous computing: smart devices, environments and interaction 137 Searching • Is a problem-solving technique that systematically explores a space of problem states, – i.e., successive and alternative stages in the problem-solving process – in order to select a goal state or a chain or path – through intermediate states to achieve a goal state. • Space of alternative solutions is searched to find an answer Ubiquitous computing: smart devices, environments and interaction 138 Searching • Much of the early research for state space search was undertaken using common board games. Why? Ubiquitous computing: smart devices, environments and interaction 139 Searching: Applications Hence search problem is expressed as: • Start state • Goal state • Goal test function • Utility function Ubiquitous computing: smart devices, environments and interaction 140 Searching: Algorithms • • • • Uninformed search No hints are available about how to reduce search space Problem search space can be represented graphically Searching involves traversing graph bread-first or depthfirst and testing each node to check if it is the goal-state. • Forward-chaining: Uninformed problem space searches tend to operate in the forward direction (progression) from start state to end goal state Ubiquitous computing: smart devices, environments and interaction 141 Uninformed or Brute Search: Breadth First Search Ubiquitous computing: smart devices, environments and interaction 142 Uninformed Search: Depth 1st Search Ubiquitous computing: smart devices, environments and interaction 143 Uninformed Search Algorithm: Design Challenges • In some cases, the search algorithm must also handle the case when the IS system cannot identify its start state – E.g., a transport vehicle is lost and requires a route to a destination • Mode of the search space may not be uniform, – E.g., different nodes may have different numbers of branches. • Also problem spaces may not be single valued but may be multiple-valued. Ubiquitous computing: smart devices, environments and interaction 144 Uninformed Search Algorithm : Design Challenges • Some problems such as games can generate extremely large search spaces, – require large amounts of computation, for uninformed search techniques • Depth-first can fail. Why? – • Hence, use of variations of informed searches. Which? Ubiquitous computing: smart devices, environments and interaction 145 Informed search techniques • Informed search is a general solution to reduce the computation of an uninformed search • Uses problem specific information to limit the problem space. • Core component of an informed search is a heuristic function – a function which depends upon the current node in a problem space, – e.g., a cost function rich returns a value to reach that current node from a previous node. Ubiquitous computing: smart devices, environments and interaction 146 Informed search techniques • What if nodes represent physical locations, what could cost function represent? – • Heuristic function which maps each node to a value depends on information about problem. • Variation of cost function to assign a first cost to reach current node from previous link to assign a second cost to go from current node to goal node. • Cost heuristic is used by A*search Ubiquitous computing: smart devices, environments and interaction 147 Informed Search: A* Search Ubiquitous computing: smart devices, environments and interaction 148 Searching Applications: Information Retrieval • A core application for searching in general is information retrieval. • Aim to reducing the cost in terms No. of goal tests for each node • How are more efficient information searching achieved? Ubiquitous computing: smart devices, environments and interaction 149 Classical (Deterministic) Planning Major planning applications? • Modern planners can be used in embodied software robots or agents as well as for complex adaptive control in machines such as particle beam accelerators. • etc Ubiquitous computing: smart devices, environments and interaction 150 Planning • Planning involves searching for a plan and then executing the plan. • Searching for a plan uses the following: – a planning model representation; – backward chaining to determine chains of actions which lead to the goal state, – forward chaining to reach the goal state from the current state; – informed search techniques – problem decomposition Ubiquitous computing: smart devices, environments and interaction 151 Planning • The planning model represents: – – – – – States Goals Actions which transition states towards goals Chains of actions to between non-adjacent states Heuristic cost functions to allow the choice of multiple chains of actions to be constrained using some heuristics • E.g., In adaptive transport scenarios example, system must pick-up multiple passengers in multiple locations, Constraints: – By path between locations to minimise fuel consumption or – Time taken or some combination of both these. Ubiquitous computing: smart devices, environments and interaction 152 Planning • Can be modelled as a graph where the nodes represent states and the link between nodes represent actions. • Links representing actions are not labelled to identify the actions – • How to define actions? Ubiquitous computing: smart devices, environments and interaction 153 EDA Action vs. Planning Action Definitions • ???? Ubiquitous computing: smart devices, environments and interaction 154 Planning • How : to make complex planning problems more solvable? • Make use of decomposition • E.g., HTA Ubiquitous computing: smart devices, environments and interaction 155 Hierarchical Task Plan for watch AV content goal Ubiquitous computing: smart devices, environments and interaction 156 Planning: Partial Order Planning • Limitation of the forward and backward state searches? – ??? – ??? • Solution? – Partial-order-planning or POP Ubiquitous computing: smart devices, environments and interaction 157 Partial Order Plan Application: to watch AV content Ubiquitous computing: smart devices, environments and interaction 158 Planning & G-IS Design • For G-IS design, planning is used to enable a chain of actions to be selected that will achieve the goal. • Hence, any action executed is part of the plan. • For some types of system interaction, the environment events may trigger actions for which there is no current plan or goal, a situated action. • Situated action could in this case trigger a goal-based IS design to form a plan for it. Ubiquitous computing: smart devices, environments and interaction 159 Non-Deterministic Planning • However, some environments may be non deterministic and partially observable models and here classical planning will fail. Why? – • Solutions? – Use Contingency planning or conditional planning Ubiquitous computing: smart devices, environments and interaction 160 Planning Application: Location (Context) awareness Start Context start Move Forward Planned Current Context Move To Side Context Deviation Re-plan & Move forward Goal Context Planned Current Context Move To Side Context Deviation Ubiquitous computing: smart devices, environments and interaction 161 Lecture Outline Types of Intelligence and IS Model Reflex & Environment Models for IS Goal and Utility based models for IS Hybrids IS Models KB IS Models KB Management: Creation & Deployment Knowledge Representation: rules, BB, semantic • Classic Logic IS Models • Soft Computing IS Models • IS System Operations environments and interaction Ubiquitous computing: smart devices, • • • • • • • 162 Summary & Revision For each chapter • See book web-site for chapter summaries, references, resources etc. • Identify new terms & concepts • Apply new terms and concepts: define, use in old and new situations & problems • Debate problems, challenges and solutions • See Chapter exercises on web-site Ubiquitous computing: smart devices, environments and interaction 163 Exercises: Define New Concepts • AI Ubiquitous computing: smart devices, environments and interaction 164 Exercise: Applying New Concepts • What should UbiCom support AI? • etc Ubiquitous computing: smart devices, environments and interaction 165