Hyperparameter Optimization and Architecture Search

HPO and architecture search run as governed L0 campaigns whose winning configuration is a versioned, reproducible artifact — never an unrecorded manual tweak.

Tuning as an auditable campaign

Model quality depends heavily on hyperparameters and architecture, but ad-hoc tuning produces models nobody can reproduce. Kronos runs hyperparameter optimization and neural architecture search as first-class L0 campaigns: a defined search space, a fixed objective on a frozen validation split, a reproducible search algorithm, and a recorded winner with full lineage.

The objective is never accuracy alone. For models destined for the machine, the search objective is multi-term: predictive error, calibration quality, inference latency against the L1 edge budget, and robustness across the domain-randomization distribution. A model that is marginally more accurate but violates the latency budget or is poorly calibrated does not win.

Search methods and constraints

• Bayesian optimization / TPE over continuous and categorical spaces
• Successive halving / Hyperband to kill weak trials early
• Multi-objective Pareto search (error vs latency vs calibration)
• Hard constraints: edge latency, memory, quantization tolerance
• Every trial logged with seed, config, and result for lineage

def objective(trial):
    cfg   = suggest(trial, SEARCH_SPACE)
    model = train(cfg, dataset, seed=trial.seed)      # reproducible
    if latency(model, EDGE_BUDGET) > EDGE_BUDGET:
        raise Pruned                                  # infeasible
    return (rmse(model, val), ece(model, val),
            -robustness(model, randomized_envs))      # multi-objective

HPO campaigns share the L0 compute substrate with retraining and are scheduled against the same budget. The winning configuration is registered and becomes the pinned recipe for future retraining, so a scheduled retrain reproduces the tuned architecture rather than re-searching. The whole campaign, not just the winner, is retained under experiment tracking.