A Framework for Uncertain Agents Nivea de C. Ferreira∗, Michael Fisher and Wiebe van der Hoek Department of Computer Science, University of Liverpool Peach Street, Chadwick Building, L69 3BX, Liverpool - UK {niveacf, michael, wiebe}@csc.liv.ac.uk Our work focuses on the formal representation and verification of agent systems. In those, representation of information and knowledge of the agents is crucial. One of the leading logical approaches to do this is using a modal epistemic logic (cf. [Meyer and van der Hoek 95]), which has a well-defined and clear semantics, suitable for modeling various other attitudes of agents as well as time ([Rao and Georgeff 95]). It also permits logical verification tools to be used for analysing a specification. Furthermore, direct execution of logical statements provides an implementation of an agent’s behaviour, which is a means of bridging the gap between an agent’s theory and its implementation, something that is not always present in contemporary agent programming frameworks [Bradshaw et al. 99]. In real world applications however, it is not enough to just specify whether or not the agent knows something: often, an agent has to make a decision under circumstances where he is not absolutely sure about the pre-conditions for his actions. It is then necessary to have a more fine-tuned mechanism to deal with uncertain information. This allows one to express that, although the agent is not sure about a particular condition, he has a bias to assume certain states of affairs, if only to decide he has to obtain more information. Environments are usually unpredictable, and an agent may also interact with other unpredictable entities, like other agents [Halpern 03]. Therefore, it is important to have formalisms that incorporate uncertainty. It is in the context of executable frameworks for agents that deal with uncertain information in which our project is situated. In short, our aim is to develop a powerful, although simple, logical language that allow agents to represent uncertainty with respect to their beliefs. More specifically, the main idea is combining an existent temporal framework, M ETATE M, and a recently specified probabilistic language, PF KD45. Together those systems describe the framework we refer as P ROT EM. In the M ETATE M framework a representation of simple dynamic agents was considered [Fisher 95]. Later extensions include a representation of deliberation within agents, and agents that have beliefs [Fisher and Ghidini 99]. However, there was no consideration of handling uncertainty. Our main aim is extending the M ETATE M language, both in terms of its logical specification and its implementation, using an appropriate logical formalism to incorporate uncertainty (associated to an agent’s beliefs). The language we use for this purpose is PF KD45. PF KD45 [de Carvalho Ferreira et al. 04] is a complete, compact and conceptually simple logical approach to probabilistic reasoning that is useful for representing and reasoning about uncertainty within computational agents. Its basic modal operator is Px> , where the meaning of Px> ϕ is: “ϕ is believed to have a probability strictly greater than x”. A peculiar characteristic of PF KD45 is that, although its syntax allows for an operator Px> for every rational number x, in the semantics it is ∗ Author gratefully acknowledges support by the Brazilian Government under CAPES-scholarship. 1 assumed that probabilities are only taken from a finite base F = {r0 , r1 , ..., rn } ⊆ [0, 1] ∩ Q. When building models for the agent specification we only need to consider finitely many ones. It may be important to highlight that this is not a real restriction, once an agent does not need an infinite set of probability values in order to reason about uncertainty. The basic idea underlying M ETATE M is building a model for an agent’s formulae. When including beliefs, instead of generating a set of choices based only upon temporal rules, both temporal and belief rules must be considered. This leads to the construction of belief contexts and (simulated) temporal sequences. With the development of the decision procedure for the PF KD45 language, we have a computational module whose output is the set of probability values that satisfy the set of probabilistic beliefs given as input. That is, a number of possible values that can be explored when having the belief contexts built. An independent combination of those approaches creates what we have called P ROT EM. In other words, P ROT EM is a fusion of M ETATE M and PF KD45 approaches, allowing us to maintain important properties of both logics. And, finally, the combined execution procedure builds the belief contexts taking into consideration the temporal and probabilistic belief rules. At this point of our research, we have the description of P ROT EM Logic, where its properties include soundness, completeness and the finite model property. Correctness of its decision procedure is another important result. Implementation of a PF KD45 decision procedure was produced as part of this project, and combining it with the existing M ETATE M execution implementation is our next step. We can then use this framework for describing an application that combines both temporal and probabilistic beliefs follows. Although this PhD project is an ongoing research activity, consistent description of the involved systems, as well as their algorithm implementations, lead us to state that we will probably succeed in the attempt of providing a logical framework, P ROT EM, that allows the representation and implementation of uncertain agents. References [Bradshaw et al. 99] J. Bradshaw, M. Greaves, H. Holmback, T. Karygiannis, B. Silverman, N. Suri, and A. Wong. Agents for the Masses: Is it possible to make development of sophisticated agents simple enough to be practical? IEEE Intelligent Systems, 14(2): 53–63, 1999. [de Carvalho Ferreira et al. 04] N. de Carvalho Ferreira, M. Fisher, and W. van der Hoek. Practical Reasoning for Uncertain Agents. In Proc. Ninth European Conf. on Logics in Artificial Intelligence (JELIA), pages 82–94, 2004. [Fisher and Ghidini 99] M. Fisher and C. Ghidini. Programming Resource-Bounded Deliberative Agents. In Proc. International Joint Conf. on Artificial Intelligence (IJCAI), pages 200–206. Morgan Kaufmann, 1999. [Fisher 95] M. Fisher. Representing and Executing Agent-Based Systems. In M. Wooldridge and N. R. Jennings, editors, Intelligent Agents, pages 307–323. Springer-Verlag, 1995. [Halpern 03] J. Y. Halpern. Reasoning About Uncertainty. MIT Press, 2003. [Meyer and van der Hoek 95] J.-J. Ch. Meyer and W. van der Hoek. Epistemic Logic for AI and Computer Science. Cambridge University Press: Cambridge, England, 1995. [Rao and Georgeff 95] A. S. Rao and M. Georgeff. BDI Agents: from Theory to Practice. In Proc. ICMAS, pages 312–319, 1995. 2