Wayne Wobcke

The implementation of a first-order logic AGM belief revision system

Belief revision is increasingly being seen as central to a number of fundamental problems in artificial intelligence such as nonmonotonic reasoning, reasoning about action, truth maintenance and database update. The authors describe the first implementation of an AGM belief revision system. The system is based on classical first-order logic, and for any finitely representable belief state, it efficiently computes expansions, contractions and revision satisfying the AGM postulates for rational belief change. The system uses a finite base to represent a belief set, and interprets a partially specified entrenchment as representing a unique most conservative entrenchment-this is motivated by considerations of evidence and by the close connections between belief revision and nonmonotonic reasoning. The authors describe in detail the algorithms for belief change, and give some examples of the systems operation.

Agents and Multi-Agent Systems Formalisms, Methodologies, and Applications

Specifying communicative multi-agent systems.- Making it up as they go along: A theory of reactive cooperation.- Agency and the logic of ability.- On the relation between interpreted systems and Kripke models.- Disjunctive logic programming and possible model semantics.- A non-monotonic ATMS based on annotated logic programs with strong negation.- Switching between reasoning and search.- The social dimension of interactions in multiagent systems.- Generating states of joint commitment between autonomous agents.- Social co-ordination among autonomous problem-solving agents.- An algorithm for plan verification in multiple agent systems.- A framework for coordination and learning among teams of agents.- An iterated hawk-and-dove game.- A game-theoretic solution of conflicts among competitive agents.- Transformation between the EMYCIN model and the Bayesian network.- Using multi-agent approach for the design of an intelligent learning environment.- A case-based strategy for solution synthesis among cooperative expert systems.

Plans and the Revison of Intentions

Although various theories of intention have been proposed, it is not obvious whether any of them can be realized in a computer implementation. Conversely, although there are many multi-agent systems, it is not clear which theories of intention they embody. In this paper, we present an initial step towards provably realizing a theory of intention in a computer system. The theory is a simplified version of Cohen and Levesques theory based on situation semantics. We present a logic of belief and intention that is sound and complete with respect to our semantics and show that the standard logical puzzles concerning intention are handled correctly within the framework. The implementation is based on a belief revision system operating under the principle of minimal change of entrenchment. The main insight behind our approach is that persistence is not a defining property of intention, but rather is a consequence of the application of the principle of minimal change to intentions. A feature of the approach is the separation between the logic used by the agent and the dynamical properties of the agents mental states. As a result, it is possible to define a simple rational agent whose intentions persist but which does not believe that its intentions persist. We show how linear hierarchical plans specified as ordered sets of beliefs and intentions can be represented and executed by a rational agent with the use of a simple interpreter.

On the Correctness of PRS Agent Programs

Although software agents are becoming more widely used, methodology for constructing agent programs is poorly understood. In this paper, we take a step towards specifying and proving correctness for a class of agent programs based on the PRS architecture, Georgeff and Lansky [9], one of the most widely used in industrial settings. We view PRS as a simplified operating system capable of concurrently running a series of plans, each of which at any time is in a state of partial execution. The PRS system is construed as using a simplified interrupt mechanism to enable it, using information about goal priorities, to recover from various contingencies so that blocked plans can be resumed and eventually successfully completed. We develop a simple methodology for PRS program construction, then present a formalism combining dynamic logic and context-based reasoning that can be used to reason about such PRS plans.

Emergent Properties of Teams of Agents in the Tileworld

An emergent property of a system is a property of the system that is not possessed by any of its components. In this paper, we examine some emergent properties of teams of agents in the Tileworld which arise from agents communicating their intentions to other agents on the same team. As expected, teams of communicating agents outperform noncommunicating agents, and teams with two-way communication abilities outperform teams where a transaction consists of a single message only. The surprising result was that, up to a certain limit, as the size of the team increased, the individual performance of the team members actually increased, even though there were more agents in competition for the resource. This is because the increased rate of replenishment of resources due to other agents consumption more than compensates an agent for the negative effects of competition. This property depends on the effect of communication in causing the agents to spread out and avoid interfering with one another. We give a partial explanation for this phenomenon. We also investigate the effectiveness of communication between agents, and show that the utility of communication varies logarithmically with the range of communication.

Agency and the Logic of Ability

Theories of ability based on the dynamic logic of programs often presuppose that the agent has complete control over its actions to the extent that execution of the action never fails. Similarly, logical theories of ‘seeing to it that’, Belnap and Perloff (1988) and ‘bringing it about, Segerberg (1989), model the result of an action without regard to the original intention of the agent, so these logics are not of direct use to formalizing the reasoning of a planning agent which must make a judgement about the likelihood of its action succeeding. In this paper, we propose an analysis of simple ability, i.e. considering only atomic actions, which is compatible with both the present and future directed sense of intention, whilst admitting the possibility of action failure. The basic idea is that an agent has the ability to do an action A to achieve some goal G if it normally brings about G when attempting to do so by doing A. We shall assume a primitive notion of ‘attempting’ or, as Bratman (1987) calls it, endeavouring, to perform an action to achieve some goal. Thus goal-directed behaviour is central to defining ability. In the latter part of the paper, we argue that this concept is also central to defining agency. We propose that agency is best understood as self-controlled goal-directed activity, where the notion of an action being under the control of an agent is intimately tied to the agents ability to perform that action successfully under normal conditions.

The representation of plans in rational agent architectures

In a previous paper, the author presented a framework for rational agent architectures that use explicit representations of beliefs and intentions and a theory of belief and intention revision. He also motivated a specific logic of belief and intention with reference to standard puzzles from the literature. He considers the relationship between the BDI-architecture and classical planning systems such as NOAH and NONLIN. In particular, he shows that any nonlinear hierarchical plan without repeated actions can be represented using his formalism in such a way that any allowable execution sequence of the plan can be realized using a system implementing AGM theory revision. This result is surprising, because the AGM theorys use of total pre-orders on beliefs suggests that nonlinear plans cannot be represented: they can because the execution of such plans is sequential.

A conditional logic for planning and plan recognition

In much work in reasoning about action, an action is modelled as a function over world states. In this paper, we consider planning and plan recognition from the point of view of an approach to modelling actions as primitive semantic objects occurring in the situations of situation semantics. We argue that constraints as captured in a hierarchy of types of situations can form the basis of an agents reasoning about action, and that the necessary constraints can be represented using a standard hierarchy of planning schemas. We present a conditional logic of constraints and show how both planning and plan recognition can be characterized as inference in the logic. We claim that our approach to formalizing action models the practice of existing planning systems more closely than traditional approaches.<<ETX>>