AtariARI Agents: Evolving Agents that use Compact State Representations
One of the initial aims of my work was to evaluate the plausibility of using NEAT to evolve high-performing agents using compact state representations.
Before, investigating methods for learning compact state representations, I used the Atari Annotated RAM Interface (AtariARI) (Anand et al. 2019) to provide high-quality, low-dimensional state representations for a variety of games.
My preliminary findings for this work were published at GECCO 2020. An extended version of this paper is available here.