| Home > Publications > Reports > Informatics (CW) |
CW 445
Robby Goetschalckx, Jan Ramon, Maurice Bruynooghe, and Hendrik Blockeel
Using expert knowledge to construct state-action aggregations for reinforcement learning
Abstract
This paper offers an approach to the problem of large state spaces for reinforcement learning by constructing a state-action pair aggregation (treating similar state-action pairs as if they were the same) with the use of domain knowledge. Arbitrary aggregation is known to give possibly very large errors. In this paper it is shown how, by using expert knowledge, a state-action pair aggregation can be constructed with an error bound which can be arbitrarily small. An algorithm for this approach is proposed, and experimental results on a number of different domains are given. Using this approach, only a limited number of episodes and a limited amount of memory are needed for convergence to a provably approximate solution.
report.pdf (307K) / mailto: R. Goetschalckx