Synaptic Pruning with MAP-elites.

Reducing the number of parameters in deep learning models is a current challenge in machine learning. We exploit the capability of MAP-Elites of illuminating the search space in a reinforcement learning problem, confronting neural networks with a different number of connections. In this work, we focus specifically on the Open AI BipedalWalker [2], a widely employed reinforcement learning benchmark, and on a deep learning model that successfully solved that task. The resulting architectures show a reduction of the synaptic connectivity of approximately 90%, equivalent to the state-of-the-art pruning techniques usually employed in supervised or generative learning. Specifically, our approach allows us to evaluate the respective impacts of sparsity and dropout. Results show that both sparsity and dropouts of hidden units are uncorrelated to the performance in our generalization test.