From: Proceedings of the First International Conference on Multiagent Systems. Copyright © 1995, AAAI (www.aaai.org). All rights reserved. An Agent Algebra for the Formal Description Multi Agent Systems and Verification of Alexander Kick Lehrstuhl Informatik fiir Ingenieure und Naturwissenschaftler UniversitKt Karlsruhe, AmFasanengarten 5 D-76128 Karisruhe, Germany kick@ira.uka.de Multi agent systems (MASs)are usually complicated and therefore error-prone. Especially the fact that there is a lot of interaction between agents can cause serious errors. Formal description and verification can make MASssuitable for control of safety critical systems such as flight control. In contrast to verfication methods, such as the use of BDI(belief, desire, intention) logics, e.g., (Rao Georgetf 1993), which can be used to verify for instance commitment properties of rational agents, our agent algebra is intended to be used for more concrete properties such as safety or progress properties (deadlock freedom, fairness, ...) of more action-oriented agents. In MASsthere can be both positive and negative interactions among agents (yon Martial 1990). On the one hand, agents may or even have to cooperate to achieve their goals. On the other hand, the agents’ actions may collide. Process algebra, e.g., (Baeten & Weijland 1990), provides a way to describe and verify parallel systems. However, cooperation and conflicts can not be expressed. Wetherefore developed an algebra, which we call agent algebra, which focuses on interaction of agents. Thus, our agent algebra is especially suited for the description of cooperation and conflict, which is a major concern in MASs. In our agent algebra single agents can perform atomic actions. Single agents are described by sequences of and choices among complex actions. Through complex actions the requirements of an agent for combined execution of its action with actions of other agents can be described. E.g., the complexaction ({a), {b, c)) scribes that an agent (A) needs two other agents which perform atomic actions b and c, respectively, simultaneously to agent A’s action a. Agent A can only proceed by this complex action and is idle if there is no other branch of possible actions. A whole MASis the parallel composition of such agent terms, e.g., AI II A2II A3. Possible conflicts between agents can be described through an incompatibility relation N, e.g., rNal, rNa2,..., which means that action r must not occur at the same time as actions al or a2. 4$2 ICMAS-9$ The structured operational semantics of our agent algebra describes what actions agents perform as time proceeds. There are two types of transitions: stransitions and normal transitions. S-transitions may be non-complete transitions in the sense that demanded actions still need to be contributed by other agents. They represent stages in the negotiation amongagents to form coalitions for joint actions. Normaltransitions represent visible joint actions which are finally taken. The structured operational semantics thus allows the tracing of the behaviour of a MAS.The operational semantics also ensures that incompatible actions which can be sensed by the agents can not occur at the same time. Verification of a MAScan be performed by transforming agent terms using a number of algebraic laws. In this way, safety and progress properties and equivalence of MASscan be proved. Interesting points of further investigation are incorporating utility, coalition formation and negotiation algorithms into the agent algebra and verifying large systems by an automatic verification tool. Acknowledgments This work was supported by DFGVo 287/5-2. References Baeten, J. C. M., and Weijland, W. P. 1990. Process Algebra, volume 18 of Cambridge Tracts in Theoretical Computer Science. Cambridge, England: Cambridge University Press. Rao, A., and Georgeff, M. 1993. A model-theoretic approach to the verfication of situated reasoning systems. In Proceedings of the 13th International Joint Conference on Artificial Intelligence, 318-324. San Mateo, California: Morgan Kaufmann. yon Martial, F. 1990. Interactions among autonomous planning agents. In Demazeau, Y., and Miiller, J., eds., Decentralized A.I., 105-119. Amsterdam, Netherlands: North-Holland.