Classification under streaming data constraints implies that training must be performed continuously, can only access individual exemplars for a short time after they arrive, must adapt to dynamic behaviour over time, and must be able to retrieve a current classifier at any time. A coevolutionary genetic programming framework is adapted to operate in non-stationary streaming data environments. Methods to generate synthetic datasets for benchmarking streaming classification algorithms are introduced, and the proposed framework is evaluated against them. The use of Pareto archiving is evaluated as a mechanism for retaining access to a limited number of useful exemplars throughout training, and several fitness sharing heuristics for archiving are evaluated. Fitness sharing alone is found to be most effective under streams with continuous (incremental) changes, while the addition of an aging heuristic is preferred when the stream has stepwise changes. Tapped delay lines are explored as a method for explicitly incorporating sequence context in cyclical data streams, and their use in combination with the aging heuristic suggests a promising route forward. / Hyperref'd copy available at: https://web.cs.dal.ca/~atwater/
| Identifer | oai:union.ndltd.org:LACETR/oai:collectionscanada.gc.ca:NSHD.ca#10222/35358 |
| Date | 13 August 2013 |
| Creators | Atwater, Aaron |
| Source Sets | Library and Archives Canada ETDs Repository / Centre d'archives des thèses électroniques de Bibliothèque et Archives Canada |
| Language | English |
| Detected Language | English |
Page generated in 0.0016 seconds