S. Kamdem, N. Sueda, and H. Ohki (Japan)
coarse coding, linear gradient descent, sarsa, reinforcement learning.
This paper presents a method based on coarse coding to
approximate the value function of a reinforcement learning
problem over a continuous domain. The approach starts
from a blank states space and gradually populates it with
states features to build the agent’s knowledge. The critical
portions of the domain are autonomously discovered and
the resolution is adaptively increased to reﬁne the optimal
policy. Experiments conducted in two benchmark domains
show that the speed of learning of this method is competitive with the most efﬁcient representations under the widely
adopted function approximation method of tile coding.