Intelligent agents II, Agent theories, architectures, and languages. IJCAI '95 workshop (ATAL), Montréal, Canada, August 19-20, 1995. Proceedings (Q1910806)

This volume contains the revised versions of 26 papers that were first presented at the second ``Agent Theories, Architectures, and Languages'' (ATAL-95) workshop, held in conjunction with the IJCAI'95 conference in Montréal, Canada, August 1995. The contributions to these proceedings fall into seven major categories which range from theoretical issues of agent modeling over agent languages and system architectures to agent testbeds and their evaluation. The first major topic are agent theories, i.e., mostly logic-based formalisms for the explanation, analysis or specification of automated agents. The notion of ``rights'' of intelligent computer agents is discussed in terms of a theory of normative positions, a logic concerned with the motivational attitudes of rational agents (preferences, goals, commitments) is introduced, tableau-based decision procedures for propositional linear-time variants of Belief-Desire-Intention (BDI) logics are presented, an agent specification logic is developed containing modalities for knowledge, intention, and ``know-how'' in a general model of action and time, a variant of process logic is formulated which allows reasoning with failure, and a specification language for agents is proposed which, in addition to the CTL\(^*\) branching time connectives, contains modal connectives for representing knowledge (information implicit within an agent's state) and choices made by the agent (``seeing to it that (stit)''). The second major topic concerns agent capabilities in terms of specific components or functionalities. The problem of enabling an agent to learn preferences from routine interactions with its world is treated and an approach to multi-agent collaborative planning is introduced which is based on a recursive Propose-Evaluate-Modify cycle of actions in which potential conflicts can be detected and resolved. The third issue, agent modelling, describes different approaches to modelling other agents' behavior or mental states from the perspective of an individual agent. In particular, the design of an effective procedure for multi-agent coordination is discussed within a game-theoretic framework, a decision-theoretic model of recursive reasoning about other agents is contrasted with a traditional game-theoretic approach, the requirements of agent tracking (dynamic modelling of other agents' higher-level desires and intentions, based on observations of their actions) are analyzed for tasks in a real-world, dynamic multi-agent environment, and a limited rationality model of resource-bounded agent reasoning is presented which selectively prunes away less important models of other agents. The section on agent architectures focuses on the integration of different components and functionalities into a coherent control framework for an individual agent. A layered control architecture for autonomous robots is discussed, a combination of the BDI paradigm with a three-layered architecture -- with distinct layers for behaviors, local planning, and cooperative planning -- is described, and an architecture is introduced for building motivated agents and providing means for focusing the agents' reasoning attention in order to avoid wasting resources. The section on multi-agent system architectures discusses methodologies and architectures for groups or networks of agents, consisting of models for the individual agent as well as of organizational models. Papers discuss the use of shells for building agent systems (offering several reusable layers of languages and services and, thus, supporting agent-based software engineering), an architecture which integrates connectionist and symbolic subsystems, and architectures inspired by cognitive principles of adaptivity or the framework of a market economy. The section on agent languages describes work on agent-based programming languages. The spectrum of languages include ones based on typed lambda calculus, reflective languages like ReActalk, CONGOLOG, an agent-oriented logic programming language which is based on Moore's logical theory of action, and the Knowledge Query and Manipulation Language (KQML), which was developed within the External Interfaces Group of the ARPA Knowledge Sharing Effort. The final section contains examples of research aimed at providing testbeds for evaluating alternatives in agent design. This includes a testbed for distributed belief revision, the evaluation of a predictive decision making model based on Markov chain theory, which suggests that system performance can be drastically improved by allowing explicit cooperation among agents through communication, and a toolkit which allows system developers to explore different options in the design of autonomous agents. An extensive bibliography with 355 entries on agent theories, architectures and languages concludes the volume.