Emergent Tangled Graph Representations for Atari Game Playing Agents

Organizing code into coherent programs and relating different programs to each other represents an underlying requirement for scaling genetic programming to more difficult task domains. Assuming a model in which policies are defined by teams of programs, in which team and program are represented using independent populations and coevolved, has previously been shown to support the development of variable sized teams. In this work, we generalize the approach to provide a complete framework for organizing multiple teams into arbitrarily deep/wide structures through a process of continuous evolution; hereafter the Tangled Program Graph (TPG). Benchmarking is conducted using a subset of 20 games from the Arcade Learning Environment (ALE), an Atari 2600 video game emulator. The games considered here correspond to those in which deep learning was unable to reach a threshold of play consistent with that of a human. Information provided to the learning agent is limited to that which a human would experience. That is, screen capture sensory input, Atari joystick actions, and game score. The performance of the proposed approach exceeds that of deep learning in 15 of the 20 games, with 7 of the 15 also exceeding that associated with a human level of competence. Moreover, in contrast to solutions from deep learning, solutions discovered by TPG are also very ‘sparse’. Rather than assuming that all of the state space contributes to every decision, each action in TPG is resolved following execution of a subset of an individual’s graph. This results in significantly lower computational requirements for model building than presently the case for deep learning.