Frames-of-Reference-Based Learning: Overcoming Perceptual Aliasing in Multistep Decision-Making Tasks

Perceptual aliasing challenges reinforcement learning agents. They struggle to learn stable policies by failing identify and disambiguate perceptually identical states in the environment that require different actions reach a goal. As agent often has only local frame of reference, it cannot represent global environment. Frame-of-reference-based is feature vertebrate intelligence allows multiple simultaneous representations an at levels abstraction. This enables resolution patterns are made up features. The evolutionary computation technique classifier systems shown promise nested single-step domains. work uses frame-of-reference concept within system non-Markov multistep Considering aliased constituent level place them appropriately holistic-level policies. Instead enumerating huge search space, evolution empowers novel evolve fitter rules experimental results show effectively solves complex environments have been challenging artificial For example, utilizes 6.5, 3.71, 3.22 steps resolve Maze10, Littman57, Woods102, respectively.