Investigating the behaviour of Q()

摘要

There is a spectrum of methods for learning robot control. At one end there are model-free methods (eg. Q-learning, AHC), and at the other there are model-based methods, (eg. dynamic programming). The advantage of one technique is the weakness of the other. Model-based methods use experience effectively, but are computationally expensive; model-free methods are cheap computationally, but require an order of magnitude more experience. In the middle ground there are now methods like Q-DYNA and prioritised sweeping which use a learned model to speed temporal credit assignment. The optimal trade-off is dependent on the relative cost of experience and computation for particular tasks. Unfortunately we frequently do not know the cost balance for a particular task. Hence the goal of this work is to understand more about the sorts of methods that might work well on a wide variety of cost ratios, and in particular how model free methods might be extended. In this paper we examine the behaviour of one such model-free algorithm, Q(λ). This algorithm shows promise because it combines the best features of Sutton's TD(λ) algorithm (1988) with those of Watkins Q-learning (1989). Despite being model-free algorithm, it has been reported to outperform prioritised sweeping, the current best method for learning a policy and a model at the same time. Here we look at the effect on its performance of using replacing or accumulating traces, and at the problem of exploration sensitivity. (3 pages)