A Gamma-Levy Hybrid MetaHeuristic for HyperParameter Tuning of Deep Q Network

Abstract

In this study we propose a novel metaheuristic algorithm; namely “Gamma-Levy Hybrid Metaheuristic with Conditional Evolution (GLHM-CE)”. The proposed algorithm is evaluated over 28 Blackbox Problems of CEC-2013, Special Session on Real-Parameter Optimization and compared with modern metaheuristic and evolutionary algorithms like SHADE, Co-DE, and JADE. The statistical results show that GLHM-CE successfully circumvents local minimas on high dimensional blackbox functions and has a fast convergence. GLHM-CE is then used to optimize the hyperparameters of a static Deep Q Neural Network evaluated on OpenAI Gym Cartpole problem. The results evaluated over a total episodal run of 5000 shows a better stability of the DQN when the hyperparameters are optimized by GLHM-CE.