Transfer of Neuroevolved Controllers in Unstable Domains

In recent years, the evolution of artificial neural networks or neuroevolution has brought promising results in solving difficult reinforcement learning problems. But, like standard RL methods, it requires that solutions be discovered in simulation and then be transferred to the real world. To date, transfer has been studied primarily in mobile robot problems that exhibit stability. This paper presents the first study of transfer in an unstable environment. Using the double pole balancing task, we simulate the process of transferring controllers from simulation to the real-world, and show that the appropriate use of noise during evolution can improve transfer significantly by compensating for inaccuracy in the simulator.