Evolving a Mario agent using cuckoo search and softmax heuristics

This paper presents a method for evolving an agent which can successfully play a level of Super Mario Brothers as implemented on the MarioAI Benchmark. The Mario search space is extremely large, making finding reasonable solutions intractable for ordinary agents. The recently introduced evolutionary algorithm, cuckoo search is especially well suited toward searching such large spaces when it employs the use of Lévy flights. Unfortunately, these Lévy flights cannot be applied to non numerical problems such as Mario. We present a modification of the algorithm which uses the Lévy distribution to effect appropriate change in a much wider set of problems, including Mario. To further optimize the search of Mario´s problem space, a softmax heuristic is presented to focus on areas with likely solutions.