Multiobjective evolution of deep learning parameters for robot manipulator object recognition and grasping

Abstract

Deep Learning (DL) is currently very popular because of its similarity to the hierarchical architecture of human brain with multiple levels of abstraction. DL has many parameters that influence the network performance. In this paper, we introduce a multiobjective evolutionary algorithm (MOEA) to optimize the DBNN parameters subject to the error rate and the network training time as two conflicting objectives. To verify the effectiveness, the proposed method is applied to the robot object recognition and grasping task. We compare the performance of the optimized DBNN model with a) DBNN with arbitrarily selected parameters and b) Deep Belief Network-Deep Neural Network (DBN-DNN). The results show that optimized DL has a superior performance in terms of training time and recognition success rate. In addition, the optimized DBNN model is effective for real-time robotic implementations.