In what way can the use of regression trees within reinforcement learning lead to more efficient reinforcement learning?
I think I understand that you can model your reward and penalty system, used in reinforcement learning, with regression trees. But how exactly does it lead to more efficient reinforcement learning?
Many reinforcement learning solutions rely on function approximation, which usually means regression, in e.g. approximating the V and Q functions or elsewhere [0].
At the moment, neural networks are most popular for performing regression here, but there are no fundamental reasons to not apply regression trees. The most important things to take into account are that the regression target (V or Q) and the collected data depend on the policy. This should be accounted for in the solution somehow.
In order to understand your question better: what do you mean with 'more efficient'? What do you compare the efficiency of the proposed approach to?
[0] See Section II of Sutton's Introduction to Reinforcement Learning
