The Evolution of Representations in Genetic Programming Trees

Abstract

Artificially intelligent machines have to explore their environment, store information about it, and use this information to improve future decision making. As such, the quest is to either provide these systems with internal models about their environment or to imbue machines with the ability to create their own models—ideally the later. These models are mental representations of the environment, and we have previously shown that neuroevolution is a powerful method to create artificially intelligent machines (also referred to as agents) that can form said representations. Furthermore, we have shown that one can quantify representations and use that quantity to augment the performance of a genetic algorithm. Instead of just optimizing for performance, one can also positively select for agents that have better representations. The neuroevolutionary approach, that improves performance and lets these agents develop representations, works well for Markov Brains, which are a form of Cartesian Genetic Programming network. Conventional artificial neural networks and their recurrent counterparts, RNNs and LSTMs, are however primarily trained by backpropagation and not evolved, and they behave differently with respect to their ability to form representations. When evolved, RNNs and LSTMs do not form sparse and distinct representations, they “smear” the information about individual concepts of the environment over all nodes in the system. This ultimately makes these systems more brittle and less capable. The question we seek to address, now, is how can we create systems that evolve to have meaningful representations while preventing them from smearing these representations? We look at genetic programming trees as an interesting computational paradigm, as they can take a lot of information in through their various leaves, but at the same time condense that computation into a single node in the end. We hypothesize that this computational condensation could also prevent the smearing of information. Here, we explore how these tree structures evolve and form representations, and we test to what degree these systems either “smear” or condense information.