| dc.description.abstract | This thesis attempts to implement a genetic algorithm for training agents within the Atari game environments. The training is performed on hardware of a widely available character, and so the results give an indication of how well these models perform on relatively inexpensive equipment available to many people. The Atari environment Space Invaders was chosen to train and test the models in. As a baseline, a Deep Q-Network (DQN) algorithm is implemented within TensorFlow's TF-Agents framework. The DQN is a popular model that has inspired many new algorithms and is often used as a comparison to alternative approaches. An adaptive genetic algorithm called ACROMUSE was implemented and compared with the performance of the DQN within the environment. This algorithm adaptively determines crossover rates, mutation rates and tournament selection size. Using measures for diversity and fitness, two subpopulations are maintained to avoid converging toward a local optimum. Based on the results found here, the algorithm did not seem to converge or produce high-performing agents, and importantly performed worse than the DQN approach. The reasons for why this algorithm fails and why other genetic algorithms have succeeded are discussed. The large number of weight parameters present in the network seem to be a barrier to good performance. It is suggested that a parallel training approach is necessary to reach the number of agents and generations where a good solution could be found. It is also shown how the number of frames skipped in the environment had a significant impact on the performance of the baseline DQN model. | |
