Using deep neuroevolution to train reinforcement learning agents
Published:
Trained three environments (Cart pole, Mountain car and Pendulum) from OpenAI Gym solving all environments in 3-5 generations trained in a few minutes on a GPU. Implemented a Deep Genetic Algorithm (GA) to train the neural network of a reinforcement learning agent. The algorithm shows promising results for DL agent training as it converges to optimum performance in very few generations. Traditionally, Neural Networks are trained using a gradient based back propagation step. This approach evolves the weights of the network using a simple, gradientfree, population based GA. The experiments were performed on relatively simple examples which takes just a few minutes to train but the underlying algorithm is capable of running on Deep Convolutional Networks with millions of parameters as well.
