Introspective multistrategy learning: On the construction of learning strategies

摘要

A central problem in multistrategy learning systems is the selection and sequencing of machine learning algorithms for particular situations. This is typically done by the system designer who analyzes the learning task and implements the appropriate algorithm or sequence of algorithms for that task. We propose a solution to this problem which enables an AI system with a library of machine learning algorithms to select and sequence appropriate algorithms autonomously. Furthermore, instead of relying on the system designer or user to provide a learning goal or target concept to the learning system, our method enables the system to determine its learning goals based on analysis of its successes and failures at the performance task. The method involves three steps: Given a performance failure, the learner examines a trace of its reasoning prior to the failure to diagnose what went wrong (blame assignment); given the resultant explanation of the reasoning failure, the learner posts explicitly represented learning goals to change its background knowledge (deciding what to learn); and given a set of learning goals, the learner uses nonlinear planning techniques to assemble a sequence of machine learning algorithms, represented as planning operators, to achieve the learning goals (learning-strategy construction). In support of these operations, we define the types of reasoning failures, a taxonomy of failure causes, a second-order formalism to represent reasoning traces, a taxonomy of learning goals that specify desired change to the background knowledge of a system, and a declarative task-formalism representation of learning algorithms. We present the Meta-AQUA system, an implemented multistrategy learner that operates in the domain of story understanding. Extensive empirical evaluations of Meta-AQUA show that it performs significantly better in a deliberative, planful mode than in a reflexive mode in which learning goals are ablated and, furthermore, that the arbitrary ordering of learning algorithms can lead to worse performance than no learning at all. We conclude that explicit representation and sequencing of learning goals is necessary for avoiding negative interactions between learning algorithms that can lead to less effective learning.