Annealed search over directed-acyclic-graph ensembles for subject-specific motor-imagery EEG decoding

Abstract

In motor-imagery (MI) brain–computer interfaces, ensemble members and their fusion rule are usually fixed in advance. In DAG–SA, simulated annealing (SA) selects the members and per-node fusion operators of a two-level directed acyclic graph (DAG) of four to eight learners, generated by a formal graph grammar from a pool of 420 distinct Common Spatial Pattern, Common Spatio-Spectral Pattern and channel log-variance classifiers. DAG–SA is compared with single-best selection, full-pool soft voting, random search with the same sampler and budget, EEGNet and a Riemannian classifier. On BCI Competition IV Datasets 1 (four subjects) and 2a (nine subjects), all methods share the splits of 20 seeds; a repeated 5 × 5-fold protocol with inner five-fold selection covers the four Dataset 1 subjects with five seeds. DAG–SA reaches 67.2% (95% CI 65.1–69.2), 64.2% (62.3–66.1) and 68.2% (66.6–69.8), against best baselines of 70.2%, 77.2% and 71.0%, and lies 1.3–3.1 points below single-best selection and random search. Budgets of up to 1 ⁢ 0 4 evaluations raise validation accuracy by 3.7–5.2 points but test accuracy by at most 3.2 points, random search attains higher validation accuracy than SA at every budget, and validation and test accuracy of sampled topologies correlate weakly ( 𝑟 = 0 . 2 8 –0.39). The 90th percentile of sampled topologies exceeds single-best selection on test splits of 22 to 30 trials, a tail not separable from test noise, and validation does not identify these topologies. Topologies selected for a subject gain an uncertain 2.7 points (Dataset 1, mainly one subject) and 0.6 points (Dataset 2a) over those selected for other subjects.

The authors' abstract, as published at the source. Machine Learning with Applications, 2026 · DOI ↗