Dynamic Approach to Information Sampling and Learning.

摘要

Given that people can process only so many pieces of information, one key aspect of learning is learning which stimulus aspects are goal relevant in the current context. In addition to injecting noise into the decision process, gathering unnecessary information can have costs in terms of time, effort, dollars, fuel, and perhaps lives. In light of these considerations, many category learning models employ selective attention mechanisms that learn which stimulus dimensions are most critical to performance. However, attention in category learning models does not direct what is encoded, but instead establishes decision weights on stimulus dimensions. Likewise, machine learning approaches to feature selection sample features not included in the final subset, and these models do not contextually determine feature relevancy. To address these shortcomings, we develop a model that selectively encodes information during learning as a function of the learner's goals, task demands, and knowledge state. The model consists of two components that are both normative, but lead to apparent non-normative behaviors when linked. One component determines the value of potential sources of information. The value of a piece of information depends on the decision maker's goals and assumptions about (i.e., knowledge of) the world, as well as the cost of the information. The second component of the model reflects the decision maker's knowledge of the world, which is used by the first component to direct information gathering. This learning component of the model is updated by the information samples selected by the first component, completing the cycle of mutual influence. We develop sub-ideal models to both capture human performance and to inform related research in machine learning. Several existing datasets are considered and novel experiments are proposed.