| Automatic generation of programs has been the focus of research in the field of software engineering,and how to make machines understand the code and write programs automatically is also an expectation in the field of artificial intelligence.In the process of algorithm design,it is difficult to achieve reproducible construction of algorithms by considering different problem characteristics,solution goals and constraints for different problems.In the case of optimization problems,for example,which are ubiquitous in life and engineering,algorithms have to be redesigned when faced with optimization problems with similar structures but different data.By examining the similarities between such optimization problems,we can make the algorithm design process more reusable,assist people in reducing mental activities while designing algorithms that face similar problems,and improve the efficiency and quality of algorithm design.In solving combinatorial optimization problems,reinforcement learning can learn algorithmic solution strategies between the same types of problems through " interaction-trial-and-error" learning,and find a reusable "metaalgorithm" algorithm design model for these problems,effectively avoiding the problem of traditional theoretical design algorithms that require a lot of expertise for manual and repeated experiments.In this article,we construct a reusable algorithm design theory and framework for solving graph optimization problems based on the unified algorithm design approach in the PAR method,and integrate it with reinforcement learning to automate the construction of algorithmic programs.Taking the shortest path and traveling salesman problems as examples,we use the PAR-based recursive algorithm construction strategy learned by the reinforcement learning method to capture the correctness of the current algorithm and continuously adjust the algorithm construction strategy to build a suitable solution for the optimization problem.Based on this,a Radl algorithm generation system for optimization problems is developed,and the reliability of the system and the generated algorithms is analyzed theoretically.The main work of this paper is as follows,which has certain innovations:1.The construction of the algorithm program is left to the intelligent body in reinforcement learning to explore the algorithm program design strategy autonomously,and finally the automatic construction of the algorithm program for graph optimization problems using the reusable algorithm design theory and algorithm framework for graph optimization problems based on the PAR method.2.In order to achieve the goal of reusable algorithm program design and automatic program generation for graph optimization problems,which further improves the automation of algorithm program development in the PAR platform,we designed and implemented an automatic generation system for solving Radl algorithm programs for graph optimization problems based on the reinforcement learning approach. |
