- Title
- Dynamic Depth for Better Generalization in Continued Fraction Regression
- Creator
- Moscato, Pablo; Ciezak, Andrew; Noman, Nasimul
- Relation
- GECCO '23: Genetic and Evolutionary Computation Conference. Proceedings of the 2023 Genetic and Evolutionary Computation Conference (Lisbon, Portugal 15-19 July, 2023) p. 520-528
- Publisher Link
- http://dx.doi.org/10.1145/3583131.3590461
- Publisher
- Association for Computing Machinery
- Resource Type
- conference paper
- Date
- 2023
- Description
- A continued fraction expansion represents a real number as an expression obtained by iteratively extracting the largest whole number from its fractional part and inverting the remainder.Continued Fraction Regression (CFR) is a method for approximating unknown target functions from data. The key idea is representing the target function as an analytic continued fraction expansion. This is achieved through an optimization approach, which searches the set of possible fractions to find the best approximating fraction for the given data. This research investigates the relationship between truncated fraction depth, accuracy, complexity, and training time in the CFR method for challenging regression problems. Specifically, low-sample synthetic datasets with Gaussian noise are considered, which we use as a proxy for low-sample dynamical systems with underlying models obscured by measurement errors.We propose and assess the performance of three depth-regulating CFR approaches against six modern symbolic regression methods. We reinforce the strong generalization capacity of the CFR method while reducing model complexity and execution time. Our method achieves the most 1st place rankings in testing against its competitors on 21 Nguyen datasets. It is never worse than 3rd for any dataset while taking at most 4% of the training time of its closest competitor.
- Subject
- continued fraction; symbolic regression; generalization; memetic algorithm
- Identifier
- http://hdl.handle.net/1959.13/1492810
- Identifier
- uon:53421
- Identifier
- ISBN:9798400701191
- Language
- eng
- Reviewed
- Hits: 451
- Visitors: 447
- Downloads: 0
Thumbnail | File | Description | Size | Format |
---|