Augmenting neuro-evolutionary adaptation with representations does not incur a speed accuracy trade-off

Abstract

Representations, or sensor-independent internal models of the environment, are important for any type of intelligent agent to process and act in an environment. Imbuing an artificially intelligent system with such a model of the world it functions in remains a difficult problem. However, using neuro-evolution as the means to optimize such a system allows the artificial intelligence to evolve proper models of the environment. Previous work has found an informationtheoretic measure, R, which measures how much information a neural computational architecture (henceforth loosely referred to as a brain) has about its environment, and can additionally be used speed up the neuro-evolutionary process. However, it is possible that this improved evolutionary adaptation comes at a cost to the brain's ability to generalize or the brain's robustness to noise. In this paper, we show that this is not the case; to the contrary, we find an improved ability of the to evolve in noisy environments when the neuro-correlate R is used to augment evolutionary adaptation.

Using MAP-Elites to direct the evolution of desired neural characteristics

“…Furthermore, previous work shows that rewarding representations (referred to as R) explicitly (along with other aspects of fitness) in a Genetic Algorithm (GA) leads to significantly better performance because "augmenting evolution with R" encourages the evolution of AIs with robust mental models (see Schossau et al (2015) or Kirkpatrick and Hintze (2019a)). However, many open questions remain.…”