Adaptive Look-Ahead Planning

We present a new adaptive connectionist planning method. By interaction with an environment a world model is progressively constructed using the backpropagation learning algorithm. The planner constructs a look-ahead plan by iteratively using this model to predict future reinforcements. Future reinforcement is maximized to derive suboptimal plans, thus determining good actions directly from the knowledge of the model network (strategic level). This is done by gradient descent in action space. The problem of nding good initial plans is solved by the use of an \experience" network (intuition level). The appropriateness of this planning method for nding suboptimal actions in unknown environments is demonstrated with a target tracking problem.