A Blackboard System for Interpreting Agent Messages Marc Cavazza (1), Steven J. Mead (1), Alexander I. Strachan (1) and Alex Whittaker i (2) (1) University of Teesside, TS1 3BA, Middlesbrough, UK. {m.o.cavazza, steven.j.mead, a.i.strachan}@tees.ac.uk (2) Sony Computer Entertainment Europe, 37, Kentish Town Road, NW1 8NX, UK. Alex_Whittaker@scee.net Abstract We describe a prototype blackboard system for coordinating multiple agents in a real-time strategy game. Agents have limited cognitive abilities and communicate with the blackboard through structured messages. Agents sense their local environment at regular intervals and send perceptual information to the blackboard working space in the form of structured messages. The blackboard system performs high-level interpretation of these low- level percepts to generate high-level tactical and strategic information. In this paper, we describe the design of the various components of the blackboard. The system has been fully implemented and experiments have been performed with data from the Dropship TM game. Introduction Simulating large numbers of agents using limited resources is a recurring challenge in computer games. Real-time strategy (RTS) games typically involve the simulation of large numbers of Non-Player Characters (NPCs) whose behaviours must be coordinated, whether to cooperate with, or oppose, the player. In real-time strategy games, where the focus is typically on the behaviours of groups of agents rather than each agent’s individual characteristics, one approach to the agent-coordination problem is to centralise system control. This reduces overall complexity by allowing the internal structure of agents to remain simple, while the complexity resides within the centralised control system. Although this approach is not appropriate for games in which importance is placed on the uniqueness of individual agents (e.g. character-based adventure games), it is well-suited to the real-time strategy genre. Another important aspect of agent- coordination is agent communication. Research in the field of multi- agent systems has developed the concept of an agent communication language (ACL), in which agents exchange information using structured messages. One area in which this method has been employed is in the field of military simulation (for instance the Control and Command Simulation Interface Language (CCSIL) (Salisbury et al., 1995)). In this paper, we describe the use of a blackboard system to address the problem of agent-coordination using centralised control and communication through structured messages. The blackboard is used as a generic approach to processing agent communications centrally. Agents generate messages containing information about their internal states and surroundings, which are sent to the blackboard working space. These low-level messages provide the initial data for the system, and are interpreted by the system in terms of tactical concepts. This approach enables us to keep individual agent design simple and efficient without having to embed complex message-parsing procedures in the agents themselves. The framework for our experiments is a computer game currently under development at Sony Computer Entertainment Europe (Camden studio), Dropship™. Dropship (Figure 1) is a battlefield simulation game in which the human player is confronted to enemy troops on the ground, at sea and in the air, according to various missions assigned to him to complete a game level. Unlike most RTS games, the human player in Dropship TM does not directly control an entire army, but controls a single unit which forms part of a larger computer- controlled force. Rather than playing the part of a commander-in-chief, the player is one pilot on the battlefield. This introduces the requirement that the game implement allied NPCs as well as enemy NPCs. Agent co-ordination can be implemented through various techniques, however, one major constraint is that the Dropship game should comprise a large number of agents: up to 400 units simultaneously populating a given level. In addition, agents have limited decision-making Figure 1: The Dropship TM Game (Copyright  SCEE 2000)