Relational Random Forests Based on Random Relational Rules

Abstract

Appropriately designing sampling policies is highly important for obtaining better control policies in reinforcement learning. In this paper, we first show that the least-squares policy iteration (LSPI) framework allows us to employ statistical active learning methods for linear regression. Then we propose a design method of good sampling policies for efficient exploration, which is particularly useful when the sampling cost of immediate rewards is high. We demonstrate the usefulness of the proposed method, named active policy iteration (API), through simulations with a batting robot.Takayuki Akiyama, Hirotaka Hachiya, Masashi SugiyamaRandom Forests have been shown to perform very well in propositional learning. FORF is an upgrade of Random Forests for relational data. In this paper we investigate shortcomings of FORF and propose an alternative algorithm, RF, for generating Random Forests over relational data. RF employs randomly generated relational rules as fully self-contained Boolean tests inside each node in a tree and thus can be viewed as an instance of dynamic propositionalization. The implementation of RF allows for the simultaneous or parallel growth of all the branches of all the trees in the ensemble in an efficient shared, but still single-threaded way. Experiments favorably compare RF to both FORF and the combination of static propositionalization together with standard Random Forests. Various strategies for tree initialization and splitting of nodes, as well as resulting ensemble size, diversity, and computational complexity of RF are also investigated.Grant Anderson, Bernhard Pfahringer