Multi-criteria optimization of ball passing in simulated soccer

Simulated soccer offers a standard real-time environment for testing decision-making methods for multi-agent systems. One critical task is determining to which team-mate the ball should be passed and the optimal point where this ball should be sent. Early methods based on enforced learning or heuristics tried to aggregate anticipated risks and pay-offs instead of flexibly balancing them. That was because scholars were not treating ball passing as a multi-criteria optimization problem in its classic sense. We propose a set of three criteria, tactical gain and two time balances, which should be balanced while selecting an optimal point on the field for passing the ball to. One of the advantages of this set of criteria is that they are treating direct and leading passes in the same way, thus offering a unified method for ball passing. In order to make this problem tractable, the continuous decision parameter space is replaced by a finite set of N points on the XY-plane, which are carefully selected using some heuristics. This set is searched for non-dominated alternatives, of which one alternative is further selected. The selection is based on the relative importance of the criteria supplied by the developer of the soccer agent. The selection method is original and uses sequential elimination of poor alternatives with respect to one criterion. Criteria are applied randomly, with the probabilities proportional to their relative importance. Experiments have shown that the multi-criteria decision-making algorithm is superior to its heuristic-based counterpart. Furthermore, we suggest that it might be implemented in the RoboCup leagues dealing with physical robots.