Opponent Modeling: An Agent for Tactical Training Willem A. van Doesburg 1 , Annerieke Heuvelink 1,2 , and Egon L. van den Broek 2 1 Department of Training and Instruction TNO Defence, Security and Safety P.O. Box 23 3769 ZG Soesterberg, The Netherlands {vandoesburg, heuvelink}@tm.tno.nl http://www.tm.tno.nl 2 Department of Artificial Intelligence Vrije Universiteit Amsterdam De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands {heuvel, egon}@few.vu.nl http://www.few.vu.nl/~{heuvel, egon} Abstract This paper describes how BDI modeling can be exploited in the design of software agents that support naval training sessions. The architecture, specifications, and embedding of the BDI agent in a simulation environment are described. Subsequently, the agent’s functioning was evaluated in complex, real life, training situations for naval officers. 1. Introduction Decision-making in complex and dynamic multi-agent environments (e.g., military missions) requires a significant effort and has proven difficult to train. Moreover, such trainings are expensive, since for training one person multiple persons are needed to play the various other roles in the training scenario. Replacing these human agents by software agents would reduce the costs substantially. Such agents should be capable of showing human-like behavior. Therefore, they should incorporate – next to expert knowledge – cognitive characteristics that can be utilized using cognitive modeling techniques [1]. Recently, the Royal NetherLands Navy (RNLN) recognized the potential of software agents for training (future) naval officers in decision making. The RNLN is interested in the development of a multi-agent system that can train a student, where cognitive agents (instead of other persons) play the roles of team member, instructor, and enemy. This research presents TACOP: a TActical Cognitive OPponent. A training scenario was developed in close cooperation with a RNLN instructor, who then provided a set of plausible goals, strategies, and actions for an enemy in that scenario. This set was used to model TACOP. The environment in which the student interacts with the TACOP was created with VR-Forces [2], see Figure 1. 2. Cognitive Agent Architecture We decided to model TACOP as a BDI-agent, since the Beliefs-Desires-Intentions architecture is known to be suitable for the generation of autonomous reactive and proactive behavior [3]. TACOP’s beliefs define his knowledge and reasoning. Two kinds of beliefs can be distinguished; (i) simple beliefs, formed passively through sensor perceptions and (ii) complex beliefs, actively formed when the agent is in a certain state of mind (represented by its beliefs, desires, and intentions) and reasons about it. In addition, beliefs are constantly updated and deleted when necessary. The desires of the agent are formed by the agent’s goals. Two types of desires can be distinguished: static desires (i.e., always activated, primary goals; e.g., self defense) and dynamic desires (i.e., emerging with a belief; e.g., fire, which is only activated when the belief is present that the target is within range). When a desire of an agent is in focus, intentions (determined by beliefs and other intentions) will be generated. An intention is planned to enable the agent to fulfill its goal and is executed as soon as possible. Subsequently, observations or actions can be generated. Moreover, a link can be made between the actions of an agent and the external (real) world. Figure 1. Screenshot of the developed training simulation environment, showing TACOP (red) in action against the trainee (blue).