Evolving Sensors in Environments of Controlled Complexity
Filippo Menczer, Richard K. Belew · The MIT Press eBooks · 1994
Sensors represent a crucial link between the evolutionary forces shaping a species' relationship with its environment, and the individual's cognitive abilities to behave and learn. We report on experiments using a new class of latent energy (LEE) models to define environments of carefully controlled complexity which allow us to state bounds for random and optimal behaviors that are independent of strategies for achieving the behaviors. Using LEE's analytic basis for defining environments, we then use neural networks (NNets) to model individuals and a steady - state genetic algorithm to model an evolutionary process shaping the NNets, in particular their sensors. Our experiments consider two types of and sensors, and variants where the NNets are not allowed to learn, learn via error correction from internal prediction, and via reinforcement learning. We find that predictive learning, even when using a larger repertoire of the more sophisticated ambient sensors, provides no advantage over NNets unable to learn. However, reinforcement learning using a small number of crude contact sensors does provide a significant advantage. Our analysis of these results points to a trade-off between the genetic robustness of sensors and their informativeness to a learning system.