|
|
Improving the performance of Continuous Action Reinforcement Learning Automata Host Publication: EWRL 2011 Authors: A. Rodriguez, M. Gagliolo, P. Vrancx, R. Grau and A. Nowé Publication Year: 2011 Number of Pages: 12
Abstract: Learning automata (LA) are policy iteration reinforcement learners
that exhibit nice convergence properties in discrete action settings. Recently, a
novel formulation for continuous action reinforcement learning automata was
proposed (CARLA), featuring an interesting exploration strategy, which is well
suited for learning from sparse reward signals. In this paper, we propose an improvement of the algorithm, and evaluate its impact on performance comparing
with the original version, as well as with another continuous LA. The experimental evaluation is carried out both in simulation and on a real control setup, showing
a clear advantage of our method, which also performs well in distributed control
settings.
|
|
