LEARNING TO CONTROL DYNAMIC-SYSTEMS WITH AUTOMATIC QUANTIZATION

Learning to control dynamic systems with unknown models is a challengi ng research problem. However, most previous work that learns qualitati ve control rules does not construct qualitative states; a proper parti tion of continuous-state variables has to be designed by human users a nd given to the learning programs. We design a new learning method tha t learns appropriate qualitative state representation and the control rules simultaneously. Our method can aggressively partition the contin uous-state variables into finer, discrete ranges until control rules b ased on these ranges are learned. As a case study, we apply our method to the benchmark control problem of cart-pole balancing (also known a s the inverted pendulum). Experimental results show that our method no t only derives different partitions for the cart-pole systems with dif ferent parameters but also learns to control the systems for an extend ed period of time from random initial positions.