Self-adaptation of Neuroevolution Algorithms Using Reinforcement Learning

Selecting an appropriate neural architecture for a given dataset is an open problem in machine learning. Neuroevolution algorithms, such as NEAT, have shown great promise in automating this process. An extension of NEAT called EXAMM has demonstrated the capability to generate recurrent neural architectures for various time series datasets. At each iteration the evolutionary process is furthered using randomly selected mutation or crossover operations, which are chosen in accordance with pre-assigned probabilities. In this paper we present a self-adapting version of EXAMM that incorporates finite action-set learning automata (FALA), a reinforcement learning technique. FALA is used to dynamically adjust the aforementioned probabilities, thereby guiding the evolutionary process, while also significantly reducing the number of required hyperparameters. It is also demonstrated that this approach improves the performance of the generated networks with statistical significance. Furthermore, the evolution of the adapted probabilities is analyzed to gain further insight into the inner workings of EXAMM.