Strategy Gradient Algorithm for the Stem Balancing Problem

Policy Gradient is a policy-based reinforcement learning algorithm that approximates the optimal policy through a parametric function. The algorithm classifies the observations by softmax through neural network to derive the policy with reference. At the same time, the objective function is set by evaluating the reward after training. Finally, the iterative formula is derived by maximizing the value of the objective function, and the parameters are updated iteratively to approach the optimal policy. As a typical multivariate, nonlinear, strongly coupled and naturally unstable problem, the stem balancing problem is very suitable for testing the ability of the policy gradient algorithm in solving the optimal policy. In this paper, the objective function is constructed, the proof of the policy gradient theorem and the derivation of the gradient iteration formula are completed, the specific setting of the stem balancing problem in the reinforcement learning under the theoretical model is introduced in detail, and the flow of the round updating policy gradient algorithm is designed.