Evolving Neural Control Systems

• Authors: N. Saravanan;David B. Fogel
• Venue: IEEE Expert: Intelligent Systems and Their Applications
• Year: 1995

Controlling unstable nonlinear systems with neural networks can be problematic. Two examples presented here show that evolutionary programming provides a feasible method for addressing such control problems.The successful application of classic control design techniques usually requires an extensive knowledge of the system to be controlled, including an accurate model of its dynamics. In some situations, this information may be difficult or even impossible to obtain. The challenge is to control a system without a priori information about its dynamics.One way to achieve this is to evolve neural networks to control the system using only sparse feedback from the system. Such neural networks can be evolved by using evolutionary programming, a member of the class of stochastic optimization techniques commonly described as evolutionary computation. Evolutionary programming has been successfully used to optimize the performance of neural networks by evolving their weights and biases.Neural networks are usually trained by an algorithm called the generalized delta rule, which computes derivatives of the error surface with respect to the weight changes by a simple application of the chain rule called backpropagation. The evolutionary programming approach to training neural networks does not require the calculation of any derivatives, and is therefore potentially useful in problem domains where such information is unavailable or might be difficult or computationally costly to obtain. The design of neurocontrollers for dynamic, nonlinear, unstable systems without a priori knowledge of the dynamics of the system is one such domain. It is this class of problems to which the current study applies evolutionary programming.More specifically, the problems we consider here involve the balancing of two poles on a moving cart, with the poles being either separated or jointed. The objective is to keep the poles upright and the cart within specified limits by simply pushing or pulling the cart. Evolutionary programming trains a feedforward neural network using only sparse feedback from the environment concerning its performance. By sparse we mean that the only feedback information available is the failure signal when either of the poles exceeds a maximum angle of deflection or when the cart reaches the end of its track.