Research On Multi-task Optimization Method Based On Knowledge Transfer

Multi-task optimization(MTO)is one of the current hot topics in the field of intelligent optimization.Multi-task optimization algorithms improve the optimization effect of each task by transferring effective knowledge between tasks when optimizing multiple tasks related to each other.However,traditional multitask optimization algorithms usually optimize small-scale task groups.When facing optimization scenarios with large-scale tasks,problems such as poor transfer effects,higher computational costs,and slower task convergence can occur.To solve these problems,researchers have proposed the multi-task quality diversity algorithm,which can generate a large number of diverse and well-performed solutions and can solve large-scale tasks in some specific scenarios,but it still suffers from the shortcomings of large randomness of knowledge transfer,single transfer mode,and easy to fall into local optimality.Therefore,this paper proposes two new multitask optimization algorithm based on the multitask quality diversity algorithm,which improves the existing multitask quality diversity algorithm.The first algorithm first designs an efficient way of grouping tasks.The algorithm calculates the degree of similarity between tasks based on their representation features,and then groups all tasks according to the degree of similarity.The tasks within a group constitute a knowledge transfer area and make the next grouped knowledge transfer process can occur only within that area.In this way,knowledge transfer occurs between similar tasks as much as possible,which reduces the probability of negative transfer occurrence and improves the efficiency of effective knowledge transfer.Next,a new knowledge transfer strategy is designed.Unlike the traditional multi-task optimization algorithm that transfers only the knowledge of the best-performing task,this algorithm records knowledge from multiple sources in each knowledge transfer area,including the performance-improving task,the best-performing task,and the information before the task change.Combining these three types of information to assist the tasks to be optimized improves their performance.At the same time,a stochastic process is introduced to give individuals the opportunity to make random variations and alleviate the problem that individuals are easily trapped in local optima.The second algorithm is an improvement on the first algorithm,applying the idea of adaptation to the first algorithm.First,an adaptive approach is used to select knowledge transfer strategies.To further improve the performance of the algorithm,four knowledge transfer strategies are designed in this algorithm,each of which is biased to transfer knowledge of a certain channel.During the evaluation process,the algorithm can adaptively select the most suitable strategy for knowledge transfer based on the performance of the strategies.Next,an adaptive approach is used to select the crossover operator.In the process of generating children,this algorithm not only follows the crossover operator in the original algorithm,but also introduces two other popular crossover operators.In the evaluation,the algorithm can adaptively select the best performing crossover operator to generate children based on the performance of each crossover operator.In this paper,the effectiveness of the algorithm is verified using a robotic arm model(representing a simple task)and a six-legged robot model(representing a difficult task).The effectiveness of the improvement is first demonstrated by ablation experiments,and then the algorithm is compared with other algorithms on two models to demonstrate the better optimization capability of the algorithm proposed in this paper.