First-Order Regression Tree System

From GM-RKB
Jump to navigation Jump to search

A First-Order Regression Tree System is a Binary Decision Tree System that implements a First-Order Regression Tree Algorithm to solve a First-Order Regression Tree Task.

See: First-Order Rule; Inductive Logic Programming; Relational Reinforcement Learning.



References

2017

An extra constraint placed on the first-order literals that are used as tests in internal nodes is that a variable that is introduced in a node (i.e., it does not occur in higher nodes) does not occur in the right subtree of the node.

Figure 1 gives an example of a first-order regression tree. The test in a node should be read as the existentially quantified conjunction of all literals in the nodes in the path from the root of the tree to that node. In the left subtree of a node, the test of the node is added to the conjunction, for the right subtree, the negation of the test should be added. For the example state description of Fig. 2, the tree would predict a Qvalue = 0. 9, since there exists no block that is both on the floor and clear, but there is a block which is on the floor and has another block on top of it. To see this, substitute BlockA in the tree with 2 (or 4) and BlockB with 1 (or 4).

The constraint on the use of variables stems from the fact that variables in the tests of internal nodes are existentially quantified. Suppose a node introduces a new variable X. Where the left subtree of a node corresponds to the fact that a substitution for X has been found to make the conjunction true, the right side corresponds to the situation where no substitution for X exists, i.e., there is no such X. Therefore, it makes no sense to refer to X in the right subtree.