The concept of an agent has become important in both Artificial Intelligence (AI) and mainstream computer science. Our aim in this paper is to point the reader at what we perceive to be the most important theoretical and practical issues associated with the design and construction of intelligent agents. For convenience, we divide these issues into three areas (though as the reader will see, the divisions are at times somewhat arbitrary). Agent theory is concerned with the question of what an agent is, and the use of mathematical formalisms for representing and reasoning about the properties of agents. Agent architectures can be thought of as software engineering models of agents; researchers in this area are primarily concerned with the problem of designing software or hardware systems that will satisfy the prop-erties specified by agent theorists. Finally, agent languages are software systems for programming and experimenting with agents; these languages may embody principles proposed by theorists. The paper is not intended to serve as a tutorial introduction to all the issues mentioned; we hope instead simply to identify the most important issues, and point to work that elaborates on them. The article includes a short review of current and potential applications of agent technology.

In this paper, we bring techniques from operations research to bear on the problem of choosing optimal actions in partially observable stochastic domains. We begin by introducing the theory of Markov decision processes (mdps) and partially observable mdps (pomdps). We then outline a novel algorithm for solving pomdps offline and show how, in some cases, a finite-memory controller can be extracted from the solution to a pomdp. We conclude with a discussion of how our approach relates to previous work, the complexity of finding exact solutions to pomdps, and of some possibilities for finding approximate solutions.

The concept of an agent has recently become important in Artificial Intelligence (AI), and its relatively youthful subfield, Distributed AI (DAI). Our aim in this paper is to point the reader at what we perceive to be the most important theoretical and practical issues associated with the design and construction of intelligent agents. For convenience, we divide the area into three themes (though as the reader will see, these divisions are at times somewhat arbitrary). Agent theory is concerned with the question of what an agent is, and the use of mathematical formalisms for representing and reasoning about the properties of agents. Agent architectures can be thought of as software engineering models of agents; researchers in this area are primarily concerned with the problem of constructing software or hardware systems that will satisfy the properties specified by agent theorists. Finally, agent languages are software systems for programming and experimenting with agents; these languages typically embody principles proposed by theorists. The paper is not intended to serve as a tutorial introduction to all the issues mentioned; we hope instead simply to identify the key issues, and point to work that elaborates on them. The paper closes with a detailed bibliography, and some bibliographical remarks. 1

An agent's ability to understand an utterance depends upon its ability to relate that utterance to the preceding discourse. The agent must determine whether the utterance begins a new segment of the discourse, completes the current segment, or contributes to it. The intentional structure of the discourse, comprised of discourse segment purposes and their interrelationships, plays a central role in this process (Grosz and Sidner, 1986). In this thesis, we provide a computational model for recognizing intentional structure and utilizing it in discourse processing. The model specifies how an agent's beliefs about the intentions underlying a discourse affects and are affected by its subsequent discourse. We characterize this process for both interpretation and generation and then provide specific algorithms for modeling the interpretation process. The collaborative planning framework of SharedPlans (Lochbaum, Grosz, and Sidner, 1990; Grosz and Kraus, 1993) provides the basis for our model ...

. The paper is concerned with the succinct axiomatization and ecient deduction of non-change, within McCarthy and Hayes' Situation Calculus. The idea behind the proposed approach is this: suppose that in a room containing a man, a robot and a cat as the only potential agents, the only action taken by the man within a certain time interval is to walk from one place to another, while the robot's only actions are to pick up a box containing the (inactive) cat and carry it from its initial place to another. We wish to prove that a certain object (such as the cat, or the doormat) did not change color. We reason that the only way it could have changed color is for the man or the robot to have painted or dyed it. But since these are not among the actions which actually occurred, the color of the object is unchanged. Thus we need no frame axioms to the eect that walking and carrying leave colors unchanged (which is in general false in multi-agent worlds), and no default schema that properties change only when we can prove they do (which is in general false in incompletely known worlds). Instead we use explanationclosure axioms specifying all primitive actions which can produce a given type of change within the setting of interest. A method similar to this has been proposed by Andrew Haas for singleagent, serial worlds. The contribution of the present paper lies in 1 showing (1) that such methods do indeed encode non-change succinctly, (2) are independently motivated, (3) can be used to justify highly ecient methods of inferring non-change, specically the \sleeping dog" strategy of STRIPS, and (4) can be extended to simple multiagent worlds with concurrent actions. An ultimate limitation may lie in the lack of a uniform strategy for deciding what ...

A fundamental assumption made by classical AI planners is that there is no uncertainty in the world: the planner has full knowledge of the conditions under which the plan will be executed and the outcome of every action is fully predictable. These planners cannot therefore construct contingency plans, i.e., plans in which different actions are performed in different circumstances. In this paper we discuss some issues that arise in the representation and construction of contingency plans and describe Cassandra, a partial-order contingency planner. Cassandra uses explicit decision-steps that enable the agent executing the plan to decide which plan branch to follow. The decision-steps in a plan result in subgoals to acquire knowledge, which are planned for in the same way as any other subgoals. Cassandra thus distinguishes the process of gathering information from the process of making decisions. The explicit representation of decisions in Cassandra allows a coherent approach to...

To build effective planning systems, it is crucial to find the right level of representation: too impoverished, and important actions and goals are impossible to express; too expressive, and planning becomes intractable.

THE aim of this thesis is to investigate logical formalisms for describing, reasoning about, specifying, and perhaps ultimately verifying the properties of systems composed of multiple intelligent computational agents. There are two obvious resources available for this task. The first is the (largely AI) tradition of reasoning about the intentional notions (belief, desire, etc.). The second is the (mainstream computer science) tradition of temporal logics for reasoning about reactive systems. Unfortunately, neither resource is ideally suited to the task: most intentional logics have little to say on the subject of agent architecture, and tend to assume that agents are perfect reasoners, whereas models of concurrent systems from mainstream computer science typically deal with the execution of individual program instructions. This thesis proposes a solution which draws upon both resources. It defines a model of agents and multi-agent systems, and then defines two execution models, which ...

We are concerned with the implications and interactions of three common expressive extensions to classical planning: conditional plans, context-dependent actions, and nondeterministic action outcomes. All of these extensions have appeared in recent work, sometimes in conjunction, but the semantics of the combination has not been fully explored. As we have argued in previous work, providing a coherent semantics for conditional planning with context-dependent actions requires that the planner's information state be modelled separately from the world state. In this paper, we present a new planning language, WCPL, encompassing these extensions. The semantics of WCPL includes an explicit treatment of the planner's information state as knowledge, as opposed to some form of context labelling. In addition to clarifying and unifying a disparate set of results from earlier work, we extend that work: WCPL handles both conditional and fail-safe plans for an action representation including both context-dependent and nondeterministic actions.

For an agent to be able to rely on a plan, he must know both that he is physically capable of carrying out the physical actions involved, and that he knows enough to carry out the plan. In this talk, we advance and discuss new definitions of "knowing enough to carry out a plan", for the case of a single agent carrying out a sequence of primitive actions one at a time. We consider both determinate and indeterminate plans. Definition: A plan P (determinate or indeterminate) is executable for agent A at time T if and only if, a. P terminates when executed starting in T ; and b. After any beginning of the execution of P starting in T , b.i A will know whether P has successfully finished; b.ii A will know of every action whether or not it is a next step of P ; and b.iii All the next steps of P are feasible. Definition: An indeterminate plan P is epistemically feasible as a task for agent A at time T if there exists a plan P 0 such that A knows in T that a. P 0 is executabl...