A0905
Title: Synergetic random forests for policy evaluation and uncertainty quantification in reinforcement learning
Authors: Ruoqing Zhu - University of Illinois at Urbana-Champaign (United States) [presenting]
Zexuan Zhang - University of Illinois at Urbana Champaign (United States)
Rui Qiu - East China Normal University (China)
Zhou Yu - East China Normal University (China)
Abstract: A new breed of random forest model is proposed, in which the splitting rule depends not only on the within-node data but also on information across the entire tree. The new model is particularly suited for situations when the splitting rule, while viewed as an estimating equation, requires further estimation of nuisance parameters that are not feasible within the node. In the proposed model, the nuisance parameter estimation is synergized across the entire tree and also progressively grows as tree nodes expand, facilitating the estimation of the main parameter of interest. A typical use case of such a model is policy evaluation in reinforcement learning when estimating the value function can utilize information from the transitional kernel. Utilizing the platform of random forests, the uncertainty of policy evaluation can also be easily quantified, which can often be challenging with other approaches.
