Real world systems have to contend with many complications which make it difficult to apply standard artificial intelligence techniques. These include an inaccessible environment, where the sensors are limited in what information they provide about the current state. Another complication is noise in the effectors of the agent. This makes the environment nondeterministic since the agent isn't sure of the effects of its actions.
Both of these complications are illustrated in the robotic soccer domain. In this domain the agents are players working cooperatively with their teammates to defeat their opponents. The agents in this domain have limited sensory information, which is given only for a limited view angle. Also, due to the nature of the game, there are unpredictable bounces and slips, which make the domain nondeterministic.
One aspect of the agent that is needed to overcome these complications in both the soccer domain and real world systems is an accurate memory model. In this paper we describe a model for the robotic soccer domain, which uses past sensory input and a probabilistic approach to maintain the model for even the inaccessible parts of the environment. Although players and the ball can move unpredictably when they are not in view, our memory model maintains reasonable estimates of their locations relative to the agent. Stationary objects, such as the goals, can be located much more precisely, even when the agent is not looking at them. Our memory model uses the locations of visible stationary objects to help determine the positions of those not in view.
The model is demonstrated using a soccer simulation program, which is explained in the following section. This is followed by a description of the memory model and the results of testing this model against a simpler one. The purpose of this paper is to allow others currently using or planning to use the Soccer Server system to duplicate our memory model for their own use.