Latent Goal Analysis for Dimension Reduction in Reinforcement Learning
; Proceedings of The 4th Workshop on Machine Learning for Interactive Systems at ICML 2015, PMLR 43:26-30, 2015.
In contrast to reinforcement learning, adaptive control formulations [Nguyen-Tuong and Peters, 2011] already come with expressive and typically low-dimensional goal and task representations, which have been generally considered more expressive than the RL setting [Kaelbling et al., 1996]. Goal and actual values in motor control define a relation similar [Rolf and Steil, 2014] to actual and target outputs in classical supervised learning settings by providing “directional information” in contrast to a mere “magnitude of an error” in reinforcement learning [Barto, 1994]. Recent work [Rolf and Asada, 2014] however showed that these two problem formulations can be transformed into each other. Hence, highly descriptive task representations can be extracted out of reinforcement learning problems by transforming them into adaptive control problems. After introducing the method called Latent Goal Analysis, we discuss the possible application of this approach as dimension reduction technique in reinforcement learning. Experimental results in a web recommender scenario confirm the potential of this technique.