The Apriori Stochastic Dependency Detection (ASDD) Algorithm for Learning Stochastic Logic Rules

Abstract

This is the unspecified version of the paper.This version of the publication may differ from the final published version.Permanent repository link: http://openaccess.city.ac.uk/3002/ Link to published version: http://dx.Abstract. Apriori Stochastic Dependency Detection (ASDD) is an algorithm for fast induction of stochastic logic rules from a database of observations made by an agent situated in an environment. ASDD is based on features of the Apriori algorithm for mining association rules in large databases of sales transactions [1] and the MSDD algorithm for discovering stochastic dependencies in multiple streams of data [15]. Once these rules have been acquired the Precedence algorithm assigns operator precedence when two or more rules matching the input data are applicable to the same output variable. These algorithms currently learn propositional rules, with future extensions aimed towards learning first-order models. We show that stochastic rules produced by this algorithm are capable of reproducing an accurate world model in a simple predator-prey environment.

Rule value reinforcement learning for cognitive agents

“…The process of building successor states described in [2], combined with the rule-value update function (2.3) allows continuous generation of next states from an initial state and updates the rule values until satisfactory values have been generated. This process is described by the following algorithm:…”

“…This is the equivalent of running Q-learning, using the rule based model for experience to build the Q(s, a) map. This method is described in [2].…”

“…This is due largely to shortcomings in the ASDD modelling method with this level of training data [2] which is reflected in the similar performance of the SR-Q results, rather than shortcomings in the RVRL algorithm. When the learned rules become a near perfect representation of the environment (at 60,000 steps training data), the results show that RVRL is capable of learning near perfect valued rules, and thus the utility of taking an action in the current state, again demonstrating that the rule values are capable of capturing a policy at least as effectively as a state action model under these conditions.…”

“…An effective method of building planning operators from experience is to use statistical significance to identify whether additional conditions are relevant to the outcome. This is the method used to create the rule sets in this work using ASDD [2]. ILP has also been used to learn rules of this form [3].…”

“…ILP has also been used to learn rules of this form [3]. Precedence between rules is used to resolve conflicts in situations where two or more rule-sets match the conditions for the same output variable (for full details of the "precedence" algorithm, see [2]). This paper is an abridged version of the work presented in [1].…”