Reinforcement learning with perceptual aliasing: the perceptual distinctions approach

摘要: It is known that Perceptual Aliasing may significantly diminish the effectiveness of reinforcement learning algorithms [Whitehead and Ballard, 1991]. aliasing occurs when multiple situations are indistinguishable from immediate perceptual input require different responses system. For example, if a robot can only see forward, yet presence battery charger behind it determines whether or not should backup, perception alone insufficient for determining most appropriate action. problematic since typically learn control policy to optimal choice action. This paper introduces predictive distinctions approach compensate caused incomplete world. An additional component, model, utilized track aspects world be visible at all times. In addition policy, model must also learned, allow stochastic actions noisy perception, probabilistic learned experience. process, system discover, on its own, important in Experimental results given simple simulated domain, issues discussed.