Evolutionary algorithms for optimising reinforcement learning policy approximation

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title
Reinforcement learning methods have become more efficient in recent years. In particular, the A3C (asynchronous advantage actor critic) approach demonstrated in Mnih et al. (2016) was able to halve the training time of the existing state-of-the-art approaches. However, these methods still require relatively large amounts of training resources due to the fundamental exploratory nature of reinforcement learning. Other machine learning approaches are able to improve the ability to train reinforcement learning agents by better processing input information to help map states to actions - convolutional and recurrent neural networks are helpful when input data is in image form that does not satisfy the Markov property. The specific required architecture of these convolutional and recurrent neural network models is not obvious given infinite possible permutations. There is very limited research giving clear guidance on neural network structure in a RL (reinforcement learning) context, and grid search-like approaches require too many resources and do not always find good optima. In order to address these, and other, challenges associated with traditional parameter optimization methods, an evolutionary approach similar to that taken by Dufourq and Bassett (2017) for image classification tasks was used to find the optimal model architecture when training an agent that learns to play Atari Pong. The approach found models that were able to train reinforcement learning agents faster, and with fewer parameters than that found by OpenAI’s model in Blackwell et al. (2018) - a superhuman level of performance.