The training set and generalization in grammatical evolution for autonomous agent navigation

Over recent years, evolutionary computation research has begun to emphasize the issue of generalization. Instead of evolving solutions that are optimized for a particular problem instance, the goal is to evolve solutions that can generalize to various different scenarios. This paper compares objective-based search and novelty search on a set of generalization oriented experiments for a navigation task using grammatical evolution (GE). In particular, this paper studies the impact that the training set has on the generalization of evolved solutions, considering: (1) the training set size; (2) the manner in which the training set is chosen (random or manual); and (3) if the training set is fixed throughout the run or dynamically changed every generation. Experimental results suggest that novelty search outperforms objective-based search in terms of evolving navigation behaviors that are able to cope with different initial conditions. The traditional objective-based search requires larger training sets and its performance degrades when the training set is not fixed. On the other hand, novelty search seems to be robust to different training sets, finding general solutions in almost all of the studied conditions with almost perfect generalization in many scenarios.