An evolutionary neural network controller for intelligent active force control

In this paper, we examine the capability of an Evolutionary Neural Network Controller (ENNC) in estimating the inertia matrix of the two-arm rigid robot. The accurate estimation of the inertia matrix is very important in the active force control loop to calculate the disturbance torques which need to be compensated in order to control a robot subjected to unknown external forces. The proposed algorithm is a modification of the EPNET algorithm proposed by Xin Yao, where we emphasise the use of crossover to explore different offsprings which do not posses strong behavioural links to their parents but still perform better than them. At the same time, the mutation operations described in EPNET, i.e. hybrid training, node deletion and node addition, are still used to maintain the behavioural link between the strong parents and their offsprings. Therefore, the introduction of the crossover will create a kind of 'survival competition' scenario between the different offsprings. In addition, this algorithm also includes the evolution of transfer (activation) functions, which play an important role in the design of the neural network. The best offspring, which represents the optimum number of nodes and types of transfer functions, is selected as the optimum neural network design for the specified problem. Then, the selected network is once again trained using back propagation with adaptive learning rate and momentum to ensure global error convergence. Finally, The fast evolutionary programming (FEP) method, which is based on the Gaussian distribution and directional mutation scheme, is incorporated to fine tune the network parameters at the end of the training session. The trained network is implemented for the active force control problem of the two-arm robot with unknown external forces. Simulation is programmed in MATLAB/SIMULINK using the Neural Network and Geatbx toolboxes. Results show significant improvement in the performance of the evolving neural network as compared to the non-evolving network.