The Application of Backpropagation Through Time and Nondominated Sorting Genetic Algorithm II in the Playing Style Imitating Process on the Computer Games of Car Racing
By : I Putu Yudi Haryasa
Faculties : Fakultas Matematika dan Ilmu Pengetahuan Alam
Department : S1 Ilmu Komputer
Imitation or modeling of playing style of play is defined as a formation of agent controller for computer games that can play the game in a manner similar to certain players. The controller is made in the form of artificial intelligence in playing a particular game that has a style of play that is similar to the human player being modeled. Problems of imitation of the behavior of the players can be seen as the sequence learning problems (in the kind of sequence prediction, which attempts to predict the next element of a sequence based on all the previous elements), so it will be used one of the standard supervised learning algorithms to train the recurrent of network namely Backpropagation Through Time (BPTT). BPTT will be combined sequentially with one of the multiobjective evolutionary algorithms, i.e. Nondominated Sorting Genetic Algorithm II (NSGA II). The implementation was conducted in the context of modeling the style of play in a car racing game of TORCS. It was conducted the evaluation of the size of the fitness by using four sizes of MSE (mean squared error) from the steering wheel commands, MSE of acceleration and braking commands, MSE of travel time, and the MSE of the damage done by a driver made in comparison to players who were imitated. The training result by using the BPTT algorithm in the best case was able to reach a value of 0.119819, 0.148762, 30.544044, and 37636. While the training results by using a combination of BPTT and NSGA II algorithm in the best case was able to reach a value of 0.120284, 0.144785, 18.731584, and 24858.77778. This indicates that the artificial intelligence based-driver as the result of a combination of BPTT and NSGA II algorithm has not been able to completely surpass the artificial driver of the algorithm results of BPTT only in the four sizes of fitness. In addition, by using a phenomenological evaluation, it was found that the development conducted by algorithms of NSGA II has not been able to provide significant visual changes to the artificial driver as a result of training by BPTT algorithm only.
Keyword : Imitation of the style of play, computer games of car racing, sequence learning, supervised learning, evolutionary multiobjective