Automated Design of Metaheuristics Using Reinforcement Learning Within a Novel General Search Framework

Automated algorithm design using proximal policy optimisation with identified features

“…DQN-based automated methods also obtained promising results for the container terminal truck routing problem and online 2D strip packing problem (Zhang et al, 2022), the vehicle routing problem and the travelling salesman problem (Dantas et al, 2021). In addition, DQN-based automated methods have also been used to select evolution operators during the search process, which is a key issue in the design of search algorithms (Yi et al, 2022) (Tian et al, 2022).…”

“…However, these policy-based RLs showed to be more stable and less prone to failure because of their stronger exploration capability. Policybased RL has been utilised to select and combine evolution operators for solving VRP in our previous study (Yi et al, 2022).…”

“…This study aims to address the key issue of automated algorithm design rather than considering the whole design space. Based on the general search framework (Yi et al, 2022), the key focus of algorithm design is on selecting and composing evolution operators at different stages during the search process.…”

“…The action space is defined by the set of evolutionary operators in Table 5, which are the algorithmic components of Evolution module in the general search framework (GSF). Refer to (Yi et al, 2022) for more information of the GSF and its basic modules. The main task of the proposed PPO method is to address the key issue of automated selection and combination of the most efficient evolutionary operators during different optimisation stages.…”

“…Based on the GCOP model, a general framework called AutoGCOP has been proposed to support the automated design of local search algorithms (Meng & Qu, 2021). A general search framework (GSF) has also been developed to support the automated design of both local search algorithms and population-based algorithms (Yi et al, 2022). Based on GSF, we further explore automated algorithm design using machine learning techniques in this study for solving the capacitated vehicle routing problem with time windows (CVRPTW).…”