Principles and Algorithms for Self-Motivated Behaviour

For planning with high uncertainty, or with too many possible end positions as in games like Go or even chess, one can almost never solve the optimal control problem and must use some receding horizon heuristics. One such heuristics is based on the idea of maximizing empowerment, namely, keep the number of possible options maximal. This has been formulated using information theoretic ideas as maximizing the information capacity between the sequence of actions and the possible state of the system at some finite horizon, but no efficient algorithm for calculating this capacity was suggested. In this work we propose a concrete and efficient way for calculating the capacity between a sequence of actions and future states, based on local linearization of the dynamics and Gaussian channel capacity calculation. I will describe the new algorithm and some of its interesting implications.