Approximate representations for multi-robot control policies that maximize mutual information