Title: What Can Agents Do Together?
A Semantics for Reasoning About Cooperative Ability

Author: Eric Werner, Ph.D.


Contact: eric.werner@cellnomica.com

Abstract

Adaptive cooperative action requires that agents be able to reason about what they can do individually (noncooperatively) and what they can do cooperatively as a group. The paper investigates the logic of noncooperative and cooperative ability. This is done by developing a formal semantics for tensed modal logics that integrate tense and modal operators with cooperative and noncooperative ability operators. The semantics incorporates formal definitions of information states and intentional states of an agent. The formal definition of group intentions, makes the semantics of group ability operators possible. The semantics can be used as a test bed for investigating the interrelationshipss between the concepts of time, information, intention, possibility, multiagent ability, and cooperation. These concepts are fundamental to distributed artificial intelligence.

1. Introduction

Flexible cooperation between agents is enhanced if agents can reason about their own and other agent's abilities. Reasoning about abilities occurs in the context of the agents own intentions as well as the intentions of other agents. For example, robot A may only be able (cooperatively) to finish his task such as the assembly of parts if the other robots supply and help assemble the parts as needed. The robot A is able to assess that he is able to perform the job only if he knows the intentional states of the other robots are such that together with his own intentions the job is achievalble. Furthermore, what actions an agent is able to perform also depends on the information the agent has about the state of the situation in which he finds himself. A semantics of assertions about abilities that supports and underlies reasoning about single and multiagent abilities will, therefore, have to include the intentional states and the information states of agents.

In this paper we investigate the relationships between information, intention and abiltiy. We do this by developing a formal semantics for the ability operator, 'can', as in 'A can \alpha', 'The group can \alpha', 'We can \alpha'. Such a semantics should be able to interpret statements like: 'If Bill has brought the tools then the team can construct the boat'. Such assertions contain both tenses and cooperative ability operators. In the case of social goals which are only achievable by more than one agent [Werner 88a], reasoning about such cooperative abilities assume a central position.

The assesment of an agent's ability over time requires the integration of several types of knowledge: state information including temporal information, and knowledge of intentions of self and others. Part of the papers contribution is that the semantics achieves this integration. Thus, it furthers our understanding of the interrelationships between an agent's knowledge and his reasoning about his ability to act.