A reinforcement learning based multi-method approach for stochastic resource constrained project scheduling problems

The Resource-Constrained Project Scheduling Problem (RCPSP) has been widely accepted as a challenging research topic due to its NP-hard nature. Because of the dynamic nature of real-world problems, stochastic-RCPSPs (SRCPSPs) are also receiving greater attention among researchers. To solve these extended RCPSPs (i.e., SRCPSPs), this paper proposes an reinforcement learning based meta-heuristic switching approach that utilizes the powers of both multi-operator differential evolution (MODE) and discrete cuckoo search (DCS) algorithms in single algorithmic framework. Reinforcement learning (RL) is introduced as a technique to select either MODE or DCS based on the diversity of population and quality of solutions. To deal with uncertain durations, a chance-constrained based approach with some belief degrees is also considered and solved by this proposed RL based multi-method approach (i.e., DECSwRL-CC). Extensive experimentation with benchmark data from the project scheduling library (PSPLIB) demonstrates the efficacy of this proposed multi-method approach. Numerous state of the art chance constrained approaches are taken from the literature to compare the proposed approach and to validate the efficacy of this multi-method approach. This particular strategy is applicable to the risk-averse decision-makers who want to realize the project schedule with a high degree of certainty.