Improving Reinforcement Learning of an Obstacle Avoidance Behavior with Forbidden Sequences of Actions

This paper is concerned with the improvement of reinforcement learning through the use of forbidden sequences of actions. A given reinforcement function can generate multiple effective behaviors. Each behavior is effective only considering the cumulative reward over time. It may not be the behavior expected by the designer. In this case, the usual solution is to modified the reinforcement function so as to introduce a new constraint related to the behavioral aspect to express. Alternatively, we propose not to modify the reinforcement function, but to add an external module containing generic forbidden sequences of actions. Experiments with the real miniature robot Khepera in a task of learning an obstacle avoidance allow to confirm the interest of this approach.