Fitness approximation for bot evolution in genetic programming

摘要: Estimating the fitness value of individuals in an evolutionary algorithm order to reduce computational expense actually calculating has been a classical pursuit practitioners. One area which could benefit from progress this endeavour is bot evolution, i.e. evolution non-playing characters computer games. Because assigning (or rather, decision tree that controls its behaviour) requires playing game, process very costly. In work, we introduce two major contributions speed up game Unreal Tournament 2004™. Firstly, method for approximation genetic programming based on idea behave similar fashion will have fitness. Thus, similarity taken at performance level, contrast commonly employed approaches are either genotypes or, less frequently, phenotypes. The performs weighted average values number individuals, attaching confidence level estimation. latter second contribution namely estimating between individuals. This involves carrying out tests consisting `static' version (with fixed inputs) with whose under evaluation and comparing results. involve limited estimation plus much lower than directly success by UT2K4 allows us expect results environments share same characteristics.