Reinforcement Learning With Its Application In Air Interception

Along with the increasing of the manoeuvrability for missiles and aircraft, the air defense guide could not be met by just using simplicity track interception guide. As a two-side conflict study method, differential games method had been applied in air interception. Whereas differential games theory was introduced from optimal control theory and precise math model was demand. Along with a nonlinear two-point boundary value problem (TPBVP) and the singular surface were met when the question was solved. So the use of differential games method was very difficult.Recently with the rising of artificial intelligence, many academicians commit themselves into introducing intelligent control theory into the study of differential games theory to settle the problem of use of it. Moreover to realize intelligent guide, the problem of automatic acquisition and use of knowledge was inevitably faced. As a way of machine learning, reinforcement learning justly could make the acquisition process of knowledge automatically. The range of knowledge could also be wide.3D air interception problem was investigated in the paper. And reinforcement learning and differential games theory was combined in the proposed method. The difficulty that brought by using traditional control theory based on the controlled object's precision math model and performance guideline to solve optimal resolution method was avoided. And according to human fuzzy thinking a group of similar air combat games rules was set up. The state was dispersed in order to minimize the action state range by using the rules. The network leaning efficiency was also increased.In traditional reinforcement learning 'Dimension Calamity' and 'Structure Credit-Assignment, problem appeared. The problem could be settled well by using Elman neural network to approximate the evaluation.The 3D air interception simulation was done. The validity of proposed method was proved by the simulation result. Using the combination of reinforcement learning and differentialgames in a high-manoeuvrability target-interception will be a further investigation topic in the future.At first the air interception's essentiality and the development of the study method were analyzed. And the based element and collection frame of design case was ordered, too. In the second chapter the characteristic, development history and groups of arithmetic of reinforcement learning were introduced. In the third chapter differential games theory based on Q-learning was designed. In the forth chapter two-side optimal method was simulated in an example of air interception. And in the fifth chapter a combination of two-side optimal and one-side optimal method was simulated in the same air interception. Both results of the two simulations were analyzed.