EARLY: Evolutionary Algorithm Optimizes Reservoir Computing Networks
The newly introduced framework, EARLY (Evolutionary Algorithm for Reservoir Learning and Yielding), enhances both the topology and hyperparameters of multi-reservoir Echo State Networks (ESNs). Drawing inspiration from the modular structure of the brain, EARLY represents architectures as graph-based genomes, utilizing crossover, mutation, and selection techniques to identify effective configurations. Its goal is to develop generic architectures that promote task generalization. When tested on temporal learning tasks from the CogScale dataset, the architectures evolved through this method surpassed those generated through random search. This innovation addresses the challenge of task-specific tuning in traditional ESNs, which frequently necessitate manual adjustments for optimal performance.
Key facts
- EARLY is a framework for evolving reservoir computing networks.
- It optimizes both topology and hyperparameters of multi-reservoir ESNs.
- Inspired by modular brain organization.
- Uses graph-based genomes with crossover, mutation, and selection.
- Goal includes creating generic architectures for task generalization.
- Tested on CogScale dataset temporal learning tasks.
- Evolved architectures outperform random search results.
- Addresses task-specific tuning limitations of classical ESNs.
Entities
Institutions
- arXiv