tailieunhanh - An Introduction to Intelligent and Autonomous Control-Chapter 12: Learning in Control

Tham khảo tài liệu 'an introduction to intelligent and autonomous control-chapter 12: learning in control', công nghệ thông tin, quản trị web phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | 12 Learning in Control Edward Grant Department of Computer Science The University of Strathclyde Livingstone Tower Richmond Street Glasgow G1 1XH . Abstract In this Chapter we review the three separate methods by which intelligent control can be applied in dynamic system control human control passive-learning and machine learning. Working from the knowledge that humans posses the ability to control complex dynamic systems by applying simple heuristics our first task was to establish those heuristics for a given dynamic domain a pole and cart system. In our work two models of the pole and cart were constructed one was a computer simulation the other a physical system. First we captured the soul of a human who was sufficiently adept and proficient at controlling the simulator. This was their rules for control. Later these same rules were encoded as a rule-base automatic controller for the purpose of conducting performance trials on the simulator and the physical system. A comparative study was also undertaken in this phase where the performance of our controller was tested against a second rule-based controller a controller using rules that were derived by interpreting the dynamic equations only. This concluded the non-learning phase. In the passive-learning phase the cause-effect signals recorded during our rule-based controller controlling the physical system were post-processed using ruleinduction to continuously refine and tune the rules needed to control the process effectively. The last phase machine-learned control assumed no a priori knowledge of the process. Here the two types of machine learning was examined the BOXES machine-learning algorithm and neural networks. In this first series of trials both performed equally well in the simulated and physical worlds. However there were certain features observed when using the machine learning algorithm in the physical domain that was particularly noteworthy. 284 INTELLIGENT AND AUTONOMOUS CONTROL 1. .