Neighbourhood approach to bisimulation in state abstraction for quantized domains

State abstraction [1] is one of solutions to the curse of dimensionality [2] problem, and possibly allows real-life application of AI algorithms. We present a new state abstraction algorithm inspired by stimulus discrimination theory from behavioral psychology [3], [4] and by current work on bisimulation theory as applied to reinforcement learning [5], [6], [7]. The new way of comparing state abstractions with the proposed notion of the ambiguity coefficient is evaluated on a well known Coffee Task domain. It is also a foundation for applying bisimulation approach to continuous domains.