Extracting refined rules from knowledge-based neural networks

Abstract

Neural networks, despite their empirically proven abilities, have been little used for the refinement of existing knowledge because this task requires a three-step process. First, knowledge must be inserted into a neural network. Second, the network must be refined. Third, the refined knowledge must be extracted from the network. We have previously described a method for the first step of this process. Standard neural learning techniques can accomplish the second step. In this article, we propose and empirically evaluate a method for the final, and possibly most difficult, step. Our method efficiently extracts symbolic rules from trained neural networks. The four major results of empirical tests of this method are that the extracted rules 1) closely reproduce the accuracy of the network from which they are extracted; 2) are superior to the rules produced by methods that directly refine symbolic rules; 3) are superior to those produced by previous techniques for extracting rules from trained neural networks; and 4) are "human comprehensible" Thus, this method demonstrates that neural networks can be used to effectively refine symbolic knowledge. Moreover, the rule-extraction technique developed herein contributes to the understanding of how symbolic and connectionist approaches to artificial intelligence can be profitably integrated.Keywords. theory refinement, integrated learning, representational shift, rule extraction from neural networks networks created using KBANN will be referred to as Knowledge-based Neural Networks--KNNs.) This step changes the representation of the rules from symbolic to neurally based, thereby making the rules refinable by standard neural learning methods.The second link of the chain is to train the KNN using a set of classified training exampies and a standard neural learning algorithm, backpropagation (Rumelhart et al., 1986) or any other method for weight optimization of feedforward neural networks. In so doing, the rules upon which the KNN are based are corrected so that they are consistent with the training examples.The final link is to extract rules from trained KNNs. This is an extremely difficult task for arbitrarily configured networks, but it is somewhat less daunting for KNNs due to properties that stem from their initial comprehensibility. Taking advantage of these properties, we developed a method to efficiently extract intelligible rules from networks. When evaluated on two real-world test problems in terms of the ability to correctly classify examples not seen during training, the method produces sets of rules that closely approximate the networks from which they came. Moreover, the extracted rules are equal, or superior, to the rules resulting from rule-refinement methods that act directly on the rules, rather than their re-representation as a neural network. We also show that our method is superior to a previous algorithm for the extraction of rules from general neural networks (e.g., Saito & Nakano, 1988;Fu, 1991).The next section contains a brief overview of our metho...

Creating advice-taking reinforcement learners

“…We plan to extend previous work on rule extraction (Craven & Shavlik, 1994;Towell & Shavlik, 1993) to produce rules in RATLE's language. We also plan to investigate the use of rule extraction as a mechanism for transfering learned knowledge between RL agents operating in the same or similar environments.…”

Handling Continuous-Valued Attributes in Decision Tree with Neural Network Modeling

“…Symbolic rules they derived from neural networks did not include input attribute relations. Also the direct conversion from neural networks to rules is related to exponential complexity when using search-based algorithm over incoming weights for each unit [5,16].…”

Untitled

“…There are many systems for theory revision in propositional and predicate logic (e.g., Koppel, Feldman, & Segre, 1994;Ourston & Mooney, 1994;Richards & Mooney, 1995;Towell & Shavlik, 1993). For instance, the EITHER system (Ourston & Mooney, 1994) uses deductive, abductive, and inductive components to revise a propositional Horn theory from a given set of counterexamples, and the KBANN system (Towell & Shavlik, 1993 uses neural networks for the same task.…”

“…1 This is the problem known in machine learning as theory revision (or knowledge-base revision) (e.g., Towell & Shavlik, 1993;Ourston & Mooney, 1994;Richards & Mooney, 1995;Koppel, Feldman, & Segre, 1994). Note that the artificial intelligence term theory, as used in this paper, has the same meaning as what the computational learning theory community calls a concept.…”