The Impact of Perceptual Aliasing on Human Learning in a Dynamic Decision Making Task

A crucial problem facing both human and artificial RL agents is correctly perceiving, and interpreting, the current state of the environment. For instance, imagine a traveler staying in an unfamiliar hotel, with each floor and exit decorated identically. Based on perceptual information alone, this guest might experience difficulty learning how to navigate towards his room, since the various hallways appear indistinguishable from one another. The problem is one of perceptual aliasing, since multiple distinct, task-relevant “states” or situations in the world map to the same percept (Whitehead & Ballard, 1991; McCallum, 1993). Note that environments may be aliased along a continuum from the perspective of the learner (e.g., only every third floor could be identical, or each floor could be perceptually distinct) and may depend on the features that the learner attends to. All else being equal, the ability of the learner should improve as the distortion induced by perceptual aliasing is reduced, and relevant states in the world uniquely map to clearly differentiated percepts.