A deep reinforcement learning based multiple meta-heuristic methods approach for resource constrained multi-project scheduling problem

Abstract

In order to solve resource-constrained multi-project scheduling problem (RCMPSP) more efficiently, this paper proposes a deep reinforcement learning algorithm based multiple meta-heuristic methods that combines the advantages of discrete cuckoo search (DCS) and particle swarm optimization (PSO). In the process of population evolution, Deep reinforcement learning was used to select the most suitable meta-heuristic algorithm (DCS and PSO) according to the diversity and quality of the population, and the reward was designed according to the evolution effect, so as to guide the algorithm to update the solutions more effectively and find the optimal solution quickly. In addition, the key steps of the CS algorithm, Levy flight and random walk, are redefined in this paper. The task movement, reverse mutation, task list recombination and adaptive and repairable swap mutation are used to make it suitable for solving discrete RCMPSP problems and improve the convergence speed of DCS algorithm. Experimental results on the latest data set (MPLIB) demonstrate the effectiveness of the proposed algorithm.