Nazal Modhej - Department of Computer Engineering, Soosangerd Branch, Islamic Azad University, Khouzestan-Iran
Jamil Neisi - Khoramshahr Branch, Islamic Azad University, Khouzestan-Iran

CMAC is a calculation model based on human cerebellum and is offered as a simple model and may be observed as a lookup table. CMAC due to high efficiency has great application in the field of modeling and control; therefore, requirement for methods to accelerate more exact learning process have made relookupers to use more diverse learning algorithms. In the present article a new algorithm for obtaining more accelerate convergence and therefore less error is offered that operates based on reinforcement learning algorithm. Whereas fixed discount rate in reinforcement learning algorithm is not suitable, a new algorithm based on discount rate of variable is offered in the present article that is applied for training CMAC. Results of simulation show that the recommended algorithm in comparison to contractual CMAC considerably decreases error.

Network, CMAC, Training Phase, System Identification, Convergence Speed, Reinforcement Learning, Discount Rate