INTELLIGENT CONTROL OF INVERTED PENDULUM THROUGH SUPERVISED LEARNING BY NEURAL NETWORK AND REINFORCEMENT LEARNING BY FUZZY Q-LEARNING

Abstract

Recently a lot of research has been done in the field of control systems to add intelligence to the system so that they can perform better control, planning and decision in real time applications. Inverted Pendulum has always been one of the preferred instruments for research work that perfectly represent a highly non linear and open loop unstable system. In this dissertation, the Googol Technology GLIP Series -2000 Inverted Pendulum has been modeled in MATLAB and its dynamical equations has been derived. The training plays a very important role in the efficiency of Artificial Neural Network (ANN).To collect the samples for training a simulink model of two Proportional- Integral-Derivative (PID) controllers in the forward path and one Linear Quadratic Regulator (LQR) in the feedback path have been made. Approximately 1000 samples were collected by adding disturbance between - -300 to ±300. To overcome the problem of slower convergence and longer training time of Back Propagation Neural Network, Levenberg-Marquardt algorithm is used for the training. The ANN controller provides better response than PID controller with lesser overshoot and lesser settling time and also automatically manages the disturbances up to -60° to +600, thus providing the intelligent control. Fuzzy Inference Systems (FIS) needs knowledge of human experts about the system to derive its rule base. Sometimes, the appropriate knowledge about the system is not available. Q-leaming is a kind of reinforcement learning that automatically tunes the rule base of the Fuzzy Logic Controller by assigning a Q-value to each rule. The technique is often called as Fuzzy Qlearning. The convergence of Q-learning depends upon the policy used for the selection of actions. In this dissertation, a new policy named as variable c-greedy policy has been proposed for Fuzzy Qlearning and compared with other policies. Two different Fuzzy Logic Controllers are designed one of 27 rules and other of 125 rules and with the help of comparison with other policies, it has been verified that variable c-greedy policy provides better convergence than other policies for both simpler 27-rule and complex 125-rule Fuzzy Logic Controller.