Autonomous Learning in a Dynamic World
Abstract
Autonomous Learning in a Dynamic World developed algorithms to autonomously adapt to concept drift that is sufficiently robust for real-world applications. The authors focused finding the 'sweet path' across a 3-dim space of forgetting-rate, bias-variance, and drift-rate. A critical finding is that drift may be domain or application dependent--thus, concept drift mapping is a requisite for adaptations. This was shown through an airlines dataset. An appropriate mapping method is then possible after characterizing the drift in the data. Using these approaches, the authors implemented this understanding to develop the novel Hoeffding Anytime Tree as and advancement over the state-of-the-art. The Hoeffding Anytime Tree splits a leaf when it has statistical evidence that the best potential split is better than no split and continually monitors the performance of the split and potential alternatives and replaces the split when it has statistical evidence that the alternative is better. This was then tested using the UCI human activity recognition data set, showing up to an order of magnitude decrease in error rate. This project resulted in 3 peer-reviewed journals, 1 conference paper, and 1 arxiv paper.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 24, 2019
- Accession Number
- AD1096539
Entities
People
- Geoff I. Webb
Organizations
- Monash University