Research On Machine Learning Algorithm And Its Application

Machine learning is the recent rise of a multi-interdisciplinary field,which involves statistics,probability theory,convex analysis,approximation theory and algorithms disciplines complexity theory.In machine learning theory,mainly related to the design and analysis that can make a computer that has the ability to self-learning algorithm.Robot Football is by far one of the applicable research topics on robot,which covers many fields such as artificial intelligence,intelligent control,graphic image processing and so on.With characteristics of real time,distribution,asynchronism and dynamic,Robot Football is a standard distributed artificial intelligence platform for research.Its simulation can not only save money,but also be convenient for people to study the properties of algorithm.Prior knowledge is unnecessary in reinforcement learning and the entity can acquire knowledge and revise behavioral strategy via interactive function with environment.These features make reinforcement learning widely used in Robot Football.On the basis of algorithm and simulation mentioned above,the thesis mainly focuses on the following aspects.Firstly,the author tries to analyze the fundamental principles of reinforcement learning,and pay more attention to the Q-Learning algorithm,the instantaneous difference algorithm and SARSA algorithm.Based on this,a modified SARSA algorithm will be put forward in combining characteristics of Robot Football.Secondly,analyze and set up MSRS platform of Robot Football simulation software;reward function and activity function are designed to reinforcement learning algorithm in simulation.Moreover,it discusses problems such as method of behavioral selection,state discretization,etc.Lastly,simulated analysis will be carried out on improved SARSA algorithm in Robot Football under single entity environment.Then the thesis makes a comparison on performance of SARSA and improved SARSA algorithm under multi-entity environment.The result shows that improved SARSA has more intelligent agent ball possession.Through Robot Football simulation platform,the thesis discusses reinforcement learning algorithm.The result turns out to be that the application of reinforcement learning can formulate the strategy of Robot Football better.