The binary-weights neural network for robot control

We propose a pure topological recurrent networks controller, which has random binary connections in hidden layer, and all hidden neurons are activated by sinusoidal functions. A direct graph encoding method and four genetic operators are implemented for using genetic programming to train this controller. Firstly, its feasibility and efficiency were validated by a pair of function approximation experiments, the results show that through evolutionary learning, this novel RNN controller can handle nonlinear problems as well as common RNN even without adjustable weights. Moreover, a simulated mobile robot was equipped with this controller, and the robot was navigated around obstacles toward a goal in physical simulation environments; during tests, this robot exhibited four successful behaviors just by topological evolving on the simple controller. This experiment reveals that this controller has the simplicity, usability and potential for robot control, it then raises the hope for further works in exploring network motifs from high level controllers.