Investigating Hyperparameter Optimization and Transferability for ES-HyperNEAT: A TPE Approach

Evolvable-Substrate HyperNEAT (ES-HyperNEAT), a bio-inspired indirect encoding that determines neuron placement and connection weights from spatial coordinates, exhibits a failure mode on MNIST as a diagnostic benchmark. Because input pixels map to a coordinate space centered at the origin, evolved networks converge on a small central cluster of input pixels, a spatial-concentration bias; prior work observed only 21% mean accuracy in this regime. Is this bias an optimization artifact or an architectural ceiling? Inspired by Mixture-of-Experts (MoE) principles, we partition the input into non-overlapping spatial segments, each assigned to a separately evolved specialist network. With 13 such experts, this design reaches 43% mean accuracy, a 106% relative improvement over the baseline. The architectural gain does not depend on data-driven aggregation: equal-weighted averaging, which uses no validation data, already yields a 70% improvement; the gain comes from partitioning, not the weighting. Receptive-field analysis shows the mechanism: partitioning forces evolution to discover features across the entire image, expanding active pixel coverage from 4% to 79%. Absolute accuracy stays below gradient-trained baselines, but the relative gain points to central bias, not the evolutionary search. Two tools are designed to generalize beyond MNIST: a receptive-field diagnostic for silent input-coverage collapse, and a spatial-partitioning remedy that restores coverage.

Neuroevolution is a representative neural architecture search paradigm that evolves both network topology and weights through evolutionary algorithms. In this paper, we propose Seq103, a unified NEAT-style neuroevolution framework for compact sequence architecture discovery. Seq103 consists of a shared evolutionary backbone and an optional recurrent extension. The shared backbone includes an elementary node-and-connection representation, per-class RMSE-based evaluation, mutation-based evolution with class-wise recombination, and elitism. The optional hidden-state mechanism extends the search space with hidden-state nodes and hidden connections, enabling temporal memory when step-wise recurrent inference is required. With this design, Seq103 applies the same core search pipeline to both step-wise recurrent and sample-wise feedforward sequence classification. In recurrent tasks, the hidden-state extension is enabled to provide temporal memory; in feedforward tasks, it is disabled while the shared evolutionary backbone remains unchanged. We evaluate Seq103 on 8 text classification datasets and the full UCRArchive2018 benchmark with 128 univariate time-series datasets. On step-wise tasks, Seq103 retains 86.96% of the best-baseline accuracy on average while using 34.6x to 3218.0x fewer parameters. On sample-wise tasks over the full UCRArchive2018 benchmark, Seq103 retains 81.95% of the best-baseline accuracy on average while using 11.8x to 160,601.0x fewer parameters.

Hyperparameter transfer allows extrapolating optimal optimization hyperparameters from small to large scales, making it critical for training large language models (LLMs). This is done either by fitting a scaling law to the hyperparameters or by a judicious choice of parameterization, such as Maximal Update ($μ$P), that renders optimal hyperparameters approximately scale invariant. In this paper, we first develop a framework to quantify hyperparameter transfer through three metrics: (1) the quality of the scaling law fit, (2) the robustness to extrapolation errors, and (3) the asymptotic loss penalty due to choice of parameterization. Next, we investigate through a comprehensive series of ablations why $μ$P appears to offer high-quality learning rate transfer relative to standard parameterization (SP), as existing theory is inadequate. We find that the overwhelming benefit of $μ$P relative to SP when training with AdamW arises simply from maximizing the learning rate of the embedding layer. In SP, the embedding layer learning rate acts as a bottleneck that induces training instabilities; increasing it by a factor of width to match $μ$P dramatically smooths out training while improving hyperparameter transfer. We also find that weight decay improves the scaling law fits, while, in the fixed token-per-parameter setting, it hurts the robustness of the extrapolation.