This paper shows the design and improvement of an autonomous agent based in using evolutionary methods to improve behavioural models (finite state machines), which are part of the individuals to evolve. This leads to the obtention of a so-called bot that follows the Gameplay track rules of the international Mario AI Championship and is able to autonomously complete different scenarios on a simulator of Super Mario Bros. game. Mono- and multi-seed approaches (evaluation in one play or multiple plays respectively) have been analysed, in order to compare respectively the performance of an approach focused in solving a specific scenario, and another more general, devoted to obtain an agent which can play successfully in different scenarios. The analysis considers the machine resources consumption, which turns in a bottleneck in some experiments. However, the methods yield agents which can finish several stages of different difficulty levels, and playing much better than an expert human player, since they can deal with very difficult situations (several enemies surrounding Mario, for instance) in real time. According to the results and considering the competition's restrictions (time limitations) and objectives (complete scenarios up to difficulty level 3), these agents have enough performance to participate in this competition track.
