Maze navigation using neural networks evolved with novelty search and differential evolution

Novelty search is a recent approach to evolving neural networks that focuses on searching for networks with new and different behaviour rather than solely focusing on finding the network with the best objective fitness. In reality the concept of novelty is short lived in the sense that nothing stays new indefinitely. Algorithms that archive the best solutions to inform the search are therefore faced with the problem that the novelty scores of these archived solutions will change from generation to generation. This research aims to address this issue by proposing two methods of adjusting novelty scores of archived solutions: 1) Novelty Decay. 2) Recalculating Archived Novelty. Novelty decay enables novelty scores to decay overtime thus enabling the search algorithm to progress while recalculating novelty scores of the archived solutions updates the novelty of these solutions at each generation. When tested on the problem of maze navigation, it is observed that novelty decay and recalculating archived novelty converge faster than both objective search and novelty search alone. Recalculating archived novelty and novelty decay perform statistically equal to one another.