Optimization And Evaluation Of Imitation Learning Algorithm Based On Gaussian Mixture Model

The research hotspot in the field of robotics is to make robots behave like humans which require the robot has the ability to learn.Imitation is a way that living things acquire skills.The introduction of imitation learning mechanism can make the robot have a certain degree of intelligence.However,imitation learning has strong dependence on demonstration information.Due to the uncertainty of demonstration quality,and the random value in the imitation learning algorithm,the effect of imitation learning is not stable.Moreover,the process of imitation learning is difficult to evaluate with a single index.Therefore,this paper focuses on the study of the optimization of robot imitation learning algorithm and the evaluation of imitation learning.The main contents are as follow.Firstly,optimization of imitation learning algorithm based on single demonstration.Gaussian mixture model(GMM)has strong reproduce ability.Therefore,GMM is widely used in the robot imitation learning.Aiming at the problem that the learning result is unstable which is due to the random selection of initial cluster value,and the problem of low learning efficiency is determined by the two steps of the characterization parameters.In this paper,an improved k-means algorithm based on the maximum and minimum distance algorithm is proposed to obtain the stable initial clustering center.Then the genetic algorithm is optimized based on Bayesian information criterion.Meanwhile,the four important parameters of GMM are obtained together.Secondly,this paper constructs a multi constraint imitation learning optimization algorithm under multiple demonstrations.Imitation learning has a strong dependence on the demonstration data,and the errors in the traditional single demonstration process will increase the difficulty of imitation learning,and causes failure.Aiming at this problem,this paper proposes an algorithm of multi constraint imitation learning based on multiple demonstrations.The probability interval is obtained as the constraint condition,and the intersection of multiple constraints is solved and the expression probability of inferior data in demonstration data is reduced.Therefore,the fluctuation of the reconstructed trajectory caused by poor data is avoided.Meanwhile,the analytic hierarchy process(AHP)is used to construct the imitative learning evaluation model.Thirdly,decision making of multi constraint learning based on Set Pair Analysis.The number of demonstration has a great influence on the learning effect of multiple constraints imitation.In order to solve this problem,this paper sets up a model of decision making with multiple constraints based on set pair analysis from the perspectives of Bayesian information criterion,computing time and goodness of fit.The model can consider both the identity and the opposition of the influencing factors,and obtain the multi constraint imitation learning scheme with high reliability.According to the experimental results,the decision rule of multi constraint imitation learning is proposed.The experimental results show that the optimization algorithm and decision model based on GMM can improve the learning accuracy and learning efficiency.Moreover,it can effectively avoid the bad effect of learning or the failure of learning,which makes the performance stable.Therefore,it has a certain guiding significance for improving the imitation learning effect under repeated demonstration.