One of the major problems of the technique that controls actions of agents by potential functions is to determine the function shapes suitable for the environment. In this paper, we study multi-agent environment where each agent autonomously acquires the parameters of potential functions. Actions of an agent are restricted by a gradient of scalar potential functions. We propose a method using reinforcement learning, especially classifier systems. Experiments with a pursuit game are done, and the following are confirmed: acquisition of the potential parameters for capturing the prey agent is possible; acquisition of the potential parameters for the two hunter agents to control each other and cooperatively capture the prey agent is also possible.