Authors:
David Bechtold
;
Alexander Wendt
and
Axel Jantsch
Affiliation:
TU Wien, Institute of Computer Technology, Gusshausstrasse 27-29, A-1040 Vienna, Austria
Keyword(s):
Reinforcement Learning, Machine Learning, Self-learning, Neural Networks, Q-learning, Deep Q-learning, Replay Memory, Artificial Intelligence, Rewards, Algorithms.
Abstract:
In recent years, interest in self-learning methods has increased significantly. A challenge is to learn to survive in a real or simulated world by solving tasks with as little prior knowledge about itself, the task, and the environment. In this paper, the state of the art methods of reinforcement learning, in particular, Q-learning, are analyzed regarding applicability to such a problem. The Q-learning algorithm is completed with replay memories and exploration functions. Several small improvements are proposed. The methods are then evaluated in two simulated environments: a discrete bit-flip and a continuous pendulum environment. The result is a lookup table of the best suitable algorithms for each type of problem.