It is difficult to achieve an optimal performance of the triple exponentially weighted moving average (TriEWMA) controller with fixed weights. To improve the performance of TriEWMA, an intelligent composite strategy using deep reinforcement learning is developed in this work. First, the weight adjustment problem is established as a Markov decision process. Then, the weights are adjusted online using the twin delayed deep deterministic policy gradient algorithm. Finally, the proposed method is applied to a chemical mechanical polishing dynamic process. Results demonstrate the effectiveness and superiority of the presented method in handling process dynamics and disturbance rejection.