Abstract
Using Dynamic Takagi-Sugeno (DTS) fuzzy models is an efficient system identification method for tasks only when the analytical model of the nonlinear real system is readily available. To overcome this drawback of the need for analytical models, a new method for system identification based on Reinforcement Learning (RL) is proposed in this study to model nonlinear systems. The learning algorithm used is semi-gradient $\mathbf{n}$-Step SARSA (State-Action-Reward-State-Action) Method which combines the Temporal-Difference (TD) learning with gradient based learning for approximating the action value function $Q\left(x_{t}, u_{t}\right)$ where $\hat{q}\left(x_{t}, u_{t}; w\right)$ is the estimate of the action-value function with parameters $w$. DTS fuzzy models with state space consequents are used as the trained models. Sum of the squared error for the output is used as the cost function for the learning. The simulation results show that the DTS fuzzy model derived based on the proposed approach approximate a dynamic nonlinear system effectively.