Research On Improved Search And Rescue Optimization Algorithm Based On Reinforcement Learning

The search and rescue optimization algorithm(SAR)is a meta-heuristic optimization algorithm proposed in 2020 for solving constrained optimization problems in engineering.As a new type of swarm intelligence optimization algorithm,SAR is flexible,simple,easy to implement and not easily trapped in local optimality.SAR consists two phase,namely social phase and individual phase.The former is used for local search,while the latter is used for global search.The two phases alternate.In this update strategy,individuals are unable to adaptively select the appropriate operator according to their own state,which affects the convergence performance of SAR.In this dissertation,the shortcomings of SAR are analyzed and the algorithm is further improved so as to extend its applications.The main work is as follows:(1)To solve the problem of slow convergence of SAR,an improved SAR named i SAR is proposed.A total of 25 functions,including single-peak function,multi-peak function,multi-modal function and composite function,are selected to test the performance of i SAR.The experimental results show that i SAR outperforms SAR,particle swarm optimization,gray wolf optimization,squirrel search and whale optimization in 25 function tests.(2)To solve the problem that SAR cannot adaptively select operators,an improved SAR named RLSAR is proposed based on reinforcement learning,which combines reinforcement learning with SAR.Using individuals of SAR as agents,a reinforcement learning model is established and the A3 C network is used for training.The training results show that the convergence speed of RLSAR is better than that of SAR.In addition,RLSAR is used to solve the rotating round table balancing pendulum problem,and the experimental results show that RLSAR can plan a better minimum envelope circle radius than SAR,which verifies the good convergence ability of RLSAR.(3)To solve the unmanned aerial vehicles path planning problem,the RLSARPP algorithm is proposed.A path adjustment phase is proposed on the basis of RLSAR.The three actions of RLSAR-PP are the path adjustment phase and the social and personal phases of SAR.RLSAR-PP is trained in randomly generated environments.The experimental results show that RLSAR-PP plans a more economical and safer path than SAR,the squirrel search algorithm and the whale optimization algorithm.The rationality of the above three action designs is verified through ablation experiments and particle execution sequence analysis,which also indicate that they have a positive impact on RLSAR-PP and enable it to effectively solve the path planning problem of unmanned aerial vehicles.