Theory, Methodology and Applications of Deep Adaptive Dynamic Programming
Abstract: Pursuing higher “intelligent” systems has become the tendency of current artificial intelligence. It is expected that the new technique is capable to perceive some complicated problems and make decisions properly. For recent years, deep learning and adaptive dynamic programming/reinforcement learning have made remarkable contribution in the field of “perception” and “decision” respectively. Deep learning has brought new techniques to perceiving high-dimensional data and processing complex information, while adaptive dynamic programming has provided advanced solutions for nonlinear system control problems. Therefore, the combination of the above two methods has been essential to advanced artificial intelligence, and has also become the research hotspot currently. Therefore, our project aims to put deep learning and adaptive dynamic programming together, to provide solutions for complicated systems which have complex information as their input. In theory and methodology, complex information will be transformed into valid character representation by deep models, and then optimal or suboptimal controllers will be computed by adaptive dynamic programming with its adaptive learning ability. Meanwhile, convergent and stable analysis will be given. In experiment and application, intelligent driving will be the main objective and auto driving simulator will be established. The proposed deep adaptive dynamic programming will be studied and compared experimentally under different scenarios, which will lay the foundation for the application of autonomous unmanned driving.
Keywords: Human-like intelligent learning control; neural network learning control; unsupervised learning control; reinforcement learning control; Intelligent learning control applications
Contact:
ZHAO Dongbin
E-mail: dongbin.zhao@ia.ac.cn
The State Key Laboratory of Management and Control for Complex Systems