A human is an expert in manipulation. We have acquired skills to perform dexterous operations based upon knowledge and experience attained over a long period of time. It is important in robotics to understand these human skills, and utilize them to bring about better robot control and operation.
It is hoped that the neural net controller can be trained and organized by simply presenting human teaching data, which implicate human intention, strategy and expertise. One weak point, however, results from this implicit method of a system construction approach. It is needed to determine the size of neural net controller. Improper size may not only incur difficulties in training neural nets, e.g. no convergence, but also cause instability and erratic behavior in machines. Therefore, it is necessary to determine the proper size of neural net controller for human control transfer.
In this paper, a new pruning method is developed, based on the penalty-term methods. This method makes the neural networks good for the generalization and reduces the retraining time after pruning weights/nodes.