Integrating Inductive Neural Network Learning and Explanation-Based Learning

TL;DR

This paper introduces Explanation-Based Neural Network Learning (EBNN), a novel method that combines inductive and analytical learning. EBNN leverages a learned domain theory (represented by neural networks) to explain observed training examples, extracting "slopes" (derivatives) that bias inductive learning. This allows EBNN to learn target functions with significantly fewer training examples and to be robust to imperfect domain knowledge.

What Problem Does This Paper Address?

Machine learning often faces a trade-off between needing vast amounts of data for purely inductive methods and requiring perfect prior knowledge for analytical methods like Explanation-Based Learning (EBL). Pure inductive learning can be data-hungry, while traditional EBL struggles with noisy or incomplete domain theories. The authors aim to bridge this gap by developing a learning system that benefits from both approaches: reducing the need for training data while remaining robust to imperfect prior knowledge. They envision a system that can operate across a spectrum of domain theory quality, from none at all to perfect.

Methods

EBNN combines inductive and analytical learning using neural networks. The core idea is to:

  1. Explain: Use a previously learned, approximate domain theory (represented by neural networks) to explain how an observed training example (e.g., a robot's action sequence leading to a goal) occurred. This explanation is a post-hoc prediction of the observed states.
  2. Analyze: Extract "slopes" (partial derivatives) of the target concept (e.g., the goal state feature) with respect to the features of the initial state and actions within the explanation. This is possible because neural networks are differentiable. These slopes capture how changes in initial states/actions affect the outcome.
  3. Refine: Use these extracted slopes, along with the observed training example's actual outcome, to update a separate neural network representing the target function. This is done both inductively (fitting the observed values) and analytically (fitting the derived slopes).

A key innovation is the LOB* (Learning from Observation, Bias by Slopes) heuristic, which estimates the accuracy of the extracted slopes based on prediction errors. This allows EBNN to down-weight the influence of slopes derived from poor domain theory explanations, making the system more robust.

Datasets

The paper evaluates EBNN on a simulated robot navigation domain. The "datasets" consist of training examples generated from this simulated environment. The domain theory itself is learned inductively from training instances, and the target function is also learned from observed examples. No standard, publicly available datasets are mentioned for evaluation.

Key Insights and Ideas

  • Slopes as Analytical Bias: Extracting derivatives (slopes) from explanations provides a powerful analytical bias for inductive learning. This bias is more informative than simple inductive biases alone, leading to faster learning.
  • Robustness through LOB*: The LOB* heuristic is crucial for making EBNN robust to imperfect domain theories. By estimating slope accuracy, it prevents noisy or incorrect explanations from derailing the learning process.
  • Graceful Degradation: EBNN doesn't fail catastrophically with imperfect domain theories; instead, it gracefully degrades to perform like a purely inductive learner when the domain theory is very poor.
  • Neural Networks as Domain Theories: Using neural networks to represent the domain theory allows for inductive learning of this theory itself, making the system more general.

Key Results

  • EBNN significantly reduces the number of training examples required to achieve a certain level of performance compared to purely inductive learning.
  • In a simulated robot control task, EBNN with a well-trained domain theory achieved comparable performance to purely inductive methods but learned much faster.
  • EBNN's performance degrades gracefully as the accuracy of the domain theory decreases, eventually matching the performance of a pure inductive system.
  • The LOB* heuristic was shown to be important for this graceful degradation; without it, performance suffered more significantly with weaker domain theories.

Limitations and Future Work

  • Propositional Domain Theory: EBNN's domain theory is propositional (represented by neural networks), unlike some symbolic EBL approaches that use first-order logic. This might limit its applicability to certain types of problems.
  • Stochastic Domains: The robustness of the LOB* heuristic in fully stochastic domains is unclear, as it relies on deterministic prediction errors.
  • Other Forms of Analytical Knowledge: Derivatives (slopes) are just one form of analytical knowledge that can be extracted. Future work could explore other types of analytical information to further accelerate learning.
  • Interaction with Pre-given Knowledge: The current EBNN learns the domain theory from scratch. Investigating how EBNN interacts with and utilizes pre-existing domain knowledge is a promising direction.
  • Scaling of Target Function Approximation: While neural networks avoid the rule-matching slowdown of symbolic systems, the accuracy of a single-network approximation for very complex target functions might degrade.

Summary updated

Abstract

Many researchers have noted the importance of combining inductive and analytical learning, yet we still lack combined learning methods that are effective in practice. We present here a learning method that combines explanation-based learning from a previously learned approximate domain theory, together with inductive learning from observations. This method, called explanation-based neural network learning (EBNN), is based on a neural network representation of domain knowledge. Explanations are constructed by chaining together inferences from multiple neural networks. In contrast with symbolic approaches to explanation-based learning which extract weakest preconditions from the explanation, EBNN extracts the derivatives of the target concept with respect to the training example features. These derivatives summarize the dependencies within the explanation, and are used to bias the inductive learning of the target concept. Experimental results on a simulated robot control task show that EBNN requires significantly fewer training examples than standard inductive learning. Furthermore, the method is shown to be robust to errors in the domain theory, operating effectively over a broad spectrum from very strong to very weak domain theories.
BibTeX
@inproceedings{1993integrating,
  title = {{Integrating Inductive Neural Network Learning and Explanation-Based Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1993},
}
Figure