SysIdBench: A Benchmark for System Identification Methods

ICLR 2026 Conference Submission18636 Authors

Keywords: system identification, benchmark, dynamic sytems

Abstract:

Modeling the behaviour of dynamic systems is a difficult problem because (i) there is a plenitude of existing system identification methods and (ii) the broadly varying characteristics of different dynamic systems are not all addressed by a single best method.
While benchmarking system identification methods has been recognized as an important asset for developers who want to select the most suitable method for their problem, these benchmarks currently lack the capabilities developers require for systematic benchmarking.
Analysing related work and our own, we have worked out five requirements on benchmarking system identification methods that have shaped the design of SysIdBench, our novel benchmark, which comprises data sets with specifically tailored data types, data splits, and evaluation metrics.
In particular, SysIdBench comprises a principle-based summarizing evaluation metrics using predictions of energy as the key measurement target, it allows for judging generalization capabilities of system identification methods, and it investigates the fulfillment of physical principles.
The code for our benchmark, including the links to the datasets, is available on [GitHub](anonymous.github.repository).

Primary Area: datasets and benchmarks

Submission Number: 18636