An ensemble-based computational approach for incremental learning in non-stationary environments related to schema- and scaffolding-based human learning

The principal dilemma in a learning process, whether human or computer, is adapting to new information, especially in cases where this new information conflicts with what was previously learned. The design of computer models for incremental learning is an emerging topic for classification and prediction of large-scale data streams undergoing change in underlying class distributions (definitions) over time; yet currently, they often ignore significant foundational learning theory that has been developed in the domain of human learning. This shortfall leads to many deficiencies in the ability to organize existing knowledge and to retain relevant knowledge for long periods of time. In this work, we introduce a unique computer-learning algorithm for incremental knowledge acquisition using an ensemble of classifiers, Learn++.NSE (Non-Stationary Environments), specifically for the case where the nature of knowledge to be learned is evolving. Learn++.NSE is a novel approach to evaluating and organizing existing knowledge (classifiers) according to the most recent data environment. Under this architecture, we address the learning problem at both the learner and supervisor end, discussing and implementing three main approaches: knowledge weighting/organization, forgetting prior knowledge, and change/drift detection. The framework is evaluated on a variety of canonical and real-world data streams (weather prediction, electricity price prediction, and spam detection). This study reveals the catastrophic effect of forgetting prior knowledge, supporting the organization technique proposed by Learn++.NSE as the most consistent performer during various drift scenarios, while also addressing the sheer difficulty in designing a system that strikes a balance between maintaining all knowledge and making decisions based only on relevant knowledge, especially in severe, unpredictable environments which are often encountered in the real-world.

Name thesis:degree_name

M.S. Engineering

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical and Computer Engineering

Year dc:date.available

2010

Author dc:creator

• Elwell, Ryan

Contributors dc:contributor

• Polikar, Robi

• Machine learning
• Electrical and Computer Engineering

Repository record dc:identifier

https://rdw.rowan.edu/etd/117

OAI identifier oai:identifier

oai:rdw.rowan.edu:etd-1116