Research On Single Machine Scheduling Algorithm Based On Adversarial Learning

Scheduling is a decision-making process to reasonably allocate resources under certain constraints.At the same time,scheduling problem,as a typical combinatorial optimization problem,has a wide application background in production,consumption,circulation service and other industries.However,it is usually difficult to get a good scheduling in the actual application process,so a fast and effective algorithm for scheduling problems has very import,ant value in many scenarios.Looking at the traditional algorithms including assignment rules,integer programming and evolutionary computation,there is no connection between the calculation process of one example and another one.Based on the idea that "similar problems should have similar solutions",the knowledge accumulated intuitively in the solution process of other examples should be beneficial to the solution of new examples,and it is also worth fully mining and utilizing in the algorithm design process.According to this understanding,this paper proposes a scheduling algorithm based on confrontation learning,and explores the design of scheduling algorithm under the framework of deep learning.In detail,the main work of this paper is as follows:From the perspective of searching for the optimal solution,considering that different examples of the same scheduling model may have different solving difficulties,at the same time,in order to train a learning model effectively,it is necessary to include data with different difficulty characteristics in the training samples as much as possible.Therefore,in order to solve the problem that it is difficult to obtain high-quality training data,this paper proposes a generator-solver framework based on adversarial learning,and unifies sample generation and training of the learning algorithms into one framework.Under the generator-solver framework based on adversarial learning,two classical single machine scheduling models with arrival time are considered and use LSTM network as the network structure of the generator-solver to design the supervised learning algorithm,in which the optimal or approximate solution of the training sample is obtained by an optimal algorithm based on dynamic programming.An alternate training method is proposed to train the generator and the solver respectively.Finally,the effectiveness of the algorithm is verified by simulation.In order to improve the quality of the solutions,and improve the generalization ability of the model,this paper use reinforcement learning method under the adversarial learning framework.The reinforcement learning method adopts Actor-Critic algorithm.The training method of the adversarial network still uses the alternate training method.The effectiveness of the algorithm is verified by simulation on the same two scheduling problem.