Reinforcement Control with Hierarchical Backpropagated Adaptive Critics

Present incremental learning methods are limited in the ability to achieve reliable credit assignment over a large number time steps (or events). However, this situation is typical for cases where the dynamical system to be controlled requires relatively frequent control updates in order to maintain stability or robustness yet has some action-consequences which must be established over relatively long periods of time. To address this problem, the learning capabilities of a control architecture comprised of two Backpropagated Adaptive Critics (BACs) in a two-level hierarchy with continuous actions are explored. The high-level BAC updates less frequently than the low-level BAC and controls the latter to some degree. The response of the low-level to high-level signals can either be determined a priori or it can emerge during learning. A general approach called Response Induction Learning is introduced to address the latter case.