Variational Reinforcement Learning for Hyper-parameter Tuning of Adaptive Evolutionary Algorithm

Researchers from Xi'an Jiaotong University have made significant contributions to the field of computational intelligence by proposing a novel approach to tuning the hyper-parameters of adaptive evolutionary algorithms. The team's variational reinforcement learning framework, named Reinforcement EM (REM), combines the expectation-maximization (EM) algorithm and a reinforcement learning algorithm to adaptively tune the hyper-parameters of differential evolution and an adaptive DE algorithm. Experimental results on the CEC 2018 test suite demonstrate that REM can achieve significantly better performance than traditional methods, including ParamILS, F-Race, and Bayesian optimization algorithm.

Key Takeaways:

• The performance of evolutionary algorithms is deeply affected by the setting of their control parameters, which should be treated as random variables.
• The proposed variational reinforcement learning framework, Reinforcement EM (REM), combines the expectation-maximization (EM) algorithm and a reinforcement learning algorithm to adaptively tune the hyper-parameters of differential evolution and an adaptive DE algorithm.
• REM can effectively adapt to new optimization problems and achieve better performance than traditional methods on the CEC 2018 test suite.
• The team proposes to use meta-learning techniques to learn good initial distributions for the hyper-parameters, which is crucial for the algorithmic performance.
• The researchers provide experimental results demonstrating the effectiveness of REM in tuning the hyper-parameters of DE and adaptive DE algorithms.
• The proposed approach has significant implications for the field of computational intelligence, particularly in the area of adaptive evolutionary algorithms.

Statistics:

• The researchers applied REM to tune the hyper-parameters of DE and adaptive DE algorithms on the CEC 2018 test suite.
• Experimental results show that REM can achieve significantly better performance than traditional methods, including ParamILS, F-Race, and Bayesian optimization algorithm.
• The study demonstrated that REM can improve the performance of DE and adaptive DE on the CEC 2018 test suite by 20% and 30%, respectively.
• The proposed approach can effectively adapt to new optimization problems, with an average improvement of 25% in performance compared to traditional methods.

Sources:

• IEEE Transactions on Emerging Topics in Computational Intelligence, 2023; 7(5): 1511-1526
• National Natural Science Foundation of China (NSFC)
• National Natural Science Foundation of China (NSFC)
• Xi'an Jiaotong University, School of Mathematics and Statistics, Natl Engn Lab Big Data Analyt, Xian 710049, People's Republic of China