关键词: 符号回归, 深度学习, 评分机制, 近端策略优化算法, 风险寻优策略梯度

Abstract: The Deep Symbolics Regression (DSR) algorithm automatically generates the expression tree from the Recurrent Neural Network (RNN) to obtain high model performance. However, it has the defects that it can not take into account the accuracy and structural simplicity of the expression tree. In this paper, a Double Decision Mechanism-based Deep Symbolic Regression (DDSR) algorithm is proposed. Based on the initial decision of RNN, the double scoring mechanism is used to comprehensively evaluate the structural simplicity and accuracy of the expression tree. Reinforcement learning is used to train the expression tree generation, which is regarded as a sequential decision-making process, and the Risk Proximal Policy Optimization algorithm is used for reward feedback to update the model parameters of the next batch. The experimental results based on public data sets show that compared with DSR algorithm, DDSR algorithm has the optimal improvement of 39.6% in the fitting performance index, and the minimum improvement is 0.5% 1%, and the overall performance was improved by 11.6%.

Key words: Symbolic regression, Deep learning, Scoring mechanism, Proximal policy optimization algorithms, Risk-seeking policy gradient