Optimizing the parameters of spiking neural networks for mobile robot implementation

Abstract

In this paper, a reward-based spike-timing-dependent plasticity method is used for the learning process of non-holonomic robots to acquire the task of target attraction. A specific fit function is developed to measure the effects of different dopamine multiplication coefficients on the training process of the spiking neural networks as well as determining the optimal operating frequencies for the network. Genetic Algorithms are used for both approaches. Several coefficients are chosen and the performance of the robot is detected based on the value of the developed fit function and the total training time. Moreover, different operational frequencies are associated with different neural regions to enhance the functionality of the network after the training phase is complete. The trained network is implemented on a mobile robot to evaluate robot performance.