Predicting the future state of the RoboCup simulation environment: heuristic and neural networks approaches

One of the challenges in multi-agent systems is the prediction of the future state of the environment. This is because the future behavior of each agent and its relationship with other agents in the environment must be predicted according to an appropriate agent model. Such a prediction about the future of the environment in a large state space, like the RoboCup simulation environment, in which most agents act on the basis of uncertain knowledge, is quite hard. Given the above context, a novel multi-agent game presentation and analysis tool has been developed. One of the agents in this tool is responsible for the prediction of individual and team behaviors. To overcome the problem of uncertain knowledge, the prediction is only made for the simulated commentator who needs to anticipate the player agent receiving the ball after a shoot by another player agent. The predictor agent has been implemented both heuristically and with neural networks. Using the logged data form RoboCup 2002 simulation league, as our test data, the neural network approach had a higher success rate of true predictions than our heuristic simulation.