The cart-pole problem is a classic problem in the theory of control of dynamic systems, often used for testing control algorithms and machine learning algorithms. This master’s thesis for the first time presents the control of a cart-pole system through the searching of continuous state space. The thesis reviews existing approaches to solving the cart-pole problem. To realize the control of the system by searching the continuous state space, the dynamic equations are derived also for driving over the uneven terrain. Various search algorithms are tested for the searching of the continuous state space, namely iterative deepening search, A* algorithm, memory-limited A* algorithm, anytime A* algorithm, and RTA*. The most suitable and frequently used algorithm was RTA*. To make the exploration of the continuous state space efficient, machine learning was used to learn a heuristic function and recognize a goal state. For this purpose, the k-nearest neighbors method and random forests were used. Controlling the system by exploring the continuous state space proved to be successful by RTA* with lookahead depth 5. In experiments with 100 randomly generated start states, the system was always successfully driven to a goal state. The control of the system was successfully tested on a flat surface as well as uneven terrain, including crossing a hill and a crater, demonstrating the robustness and effectiveness of the developed approach.
|
