LSTMを用いたDeep Neuroevolutionによるロボティックスワームの群れ行動生成

抄録

Deep Neuroevolution is a method of optimizing a Deep Neural Network (DNN) using evolutionary computation. By applying this approach for designing the robot controller, the robotic swarm is able to generate collective behavior from raw image inputs. However, we need to explore how to set up the structure of DNN and how theses settings affect the performance. In this paper, we applied DCNN and DCRNN to the controller of a robotic swarm. DCNN is a DNN that can automatically extract image features. DCRNN is a combination of DCNN and Long Short-Terms Memory network. We discussed the performance of DCNN and DCRNN in a two-target area navigation task under the environments with and without an obstacle. The results showed that DCRNN recorded better performance than DCNN in two both environments.