TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH

Introduction

• Because forecasting requires chronological validation and deep-model evaluations can be costly, TS-CASH comparisons should align temporal protocols and finite trial budgets.Existing forecasting AutoML studies differ in candidate spaces, optimized components, objectives, and evaluation protocols, confounding search strategy with study design.Source: Introduction · p. 1
• TRACE-CASH combines trial-history-conditioned grouped actor–critic candidate generation with fixed rules for model coverage, validation-guided exploitation, and renewed exploration after stalled progress.Separate model, temporal, architectural, and training actors use shared representations of previous outcomes and search progress; model-specific decoding produces the evaluated configuration.Source: Introduction · p. 1
• TRACE-CASH has the lowest mean rank on MASE and WQL, and descriptively the lowest window-averaged test-MASE rank in the predefined full and late windows.The study evaluates 41 task variants with three seeds.Source: Introduction · p. 1
• The study compares seven heterogeneous methods in a common nine-family conditional search space using the same chronological evaluator, attempted-trial budget, validity checks, and incumbent rule.The comparison retains each method’s candidate-generation mechanism.Source: Introduction · p. 1

Forecasting and HPO Benchmarks

• TRACE-CASH evaluates sequential search across heterogeneous forecasting models with a shared chronological evaluator, complementing TSBench’s offline comparison of four search methods on 20 datasets.The Monash Archive curates diverse forecasting datasets, TFB provides a unified evaluation pipeline, and YAHPO Gym offers repeatable surrogate HPO scenarios.Source: Forecasting and HPO Benchmarks · p. 2

AutoML for Time-Series Forecasting

• Time-series AutoML automates forecasting pipelines through model selection, optimization, ensembling, or meta-learning; TRACE-CASH instead searches configurations within a fixed model pool.Prior approaches include AutoAI-TS, AutoGluon–TimeSeries, auto-sktime, Auto-PyTorch-TS, AutoCTS+, AutoXPCR, and knowledge transfer from previous searches.Source: AutoML for Time-Series Forecasting · p. 2

CASH and Conditional Search Spaces

• CASH creates a hierarchical conditional space where only the selected algorithm’s hyperparameters are active; TRACE-CASH adapts task-local policies from trial history in this setting.SMAC3 uses Bayesian optimization and TPE uses Parzen density models; forecasting CASH also includes model-specific temporal and architectural choices under chronological validation.Source: CASH and Conditional Search Spaces · p. 2

Decision Decomposition in AutoML

• MA2ML and MAT-HPO divide optimization among specialized agents; TRACE-CASH groups decisions for conditional forecasting CASH, where the chosen model determines active hyperparameters.MA2ML addresses image-classification augmentation, architecture, and hyperparameters, while MAT-HPO assigns architecture, class-weight, and training decisions to module-specific Transformer agents (Wang et al. 2023; Cheng, Wu, and Tseng 2025).Source: Decision Decomposition in AutoML · p. 2

Language-Model-Assisted Search

• LLAMBO uses language models for warm starts, sampling, and surrogate modeling; LB-MCTS routes between language models and Bayesian proposers; SLLMBO combines suggestions with TPE (Liu et al. 2024; Xu et al. 2026; Mahammadli and Ertekin 2025).Source: Language-Model-Assisted Search · p. 2

• All methods use a common nine-family conditional search-space design, with model fitting, validation, and testing following a chronological protocol.Source: Problem Formulation · p. 2
• Each candidate is a model family paired with hyperparameters from that family’s domain; admissibility does not guarantee successful fitting or forecasting.Source: Problem Formulation · p. 2
• Under budget B, every attempted configuration consumes budget; valid trials must complete successfully, yield finite validation MASE, and pass shared forecast-output checks.Source: Problem Formulation · p. 2
• The incumbent is selected by validation MASE, with ties going to the earliest attempt; test scores neither rank configurations nor break ties, and WQL uses the same incumbent.Source: Problem Formulation · p. 2
• TRACE-CASH learns from immediate trial rewards, with recorded outcomes updating replay and the state conditioning later candidate generation; evaluation concerns the best incumbent within budget.Source: Problem Formulation · p. 2

Proposed Method: TRACE-CASH

• TRACE-CASH combines history-conditioned candidate generation with fixed rules for initial model coverage, outcome-guided allocation, and renewed exploration after stagnation.The rules remain fixed across task–seed runs, while learned components are reinitialized for each run.Source: Proposed Method: TRACE-CASH · p. 3

Conditional Configuration Generation

• The state summarizes recent outcomes, search progress, best validation MASE, model coverage, and model-use history, while excluding numerical test-score magnitudes.Exploration can substitute a uniformly sampled normalized action; otherwise, the policies generate the initial joint action.Source: Conditional Configuration Generation · p. 3
• A shared trial-history state feeds a categorical model actor and three model-conditioned Transformer actors for temporal, architectural, and training actions; a model-specific decoder applies relevant settings.Separating model choice from mixed-type settings avoids treating every model as having the same active hyperparameters; normalized group outputs are decoded in the evaluated model’s hyperparameter space.Source: Conditional Configuration Generation · pp. 3–4, Proposed Method: TRACE-CASH · p. 3

Actor–Critic Learning from Trial History

• Valid trials receive clipped negative validation MASE rewards and failed or invalid trials fixed penalties; rewards train the actors but do not determine incumbent eligibility or ranking.Each evaluated configuration is scored directly, and its effect on later decisions is carried by the updated history state; test-score magnitudes do not enter reward or incumbent ranking.Source: Actor–Critic Learning from Trial History · p. 4
• One critic per actor scores masked views under shared history and model context, with summed scores trained toward the shared reward; actor updates resample unevaluated actions from replayed states.Critics learn from configurations actually evaluated, while fixed rules apply only when selecting a candidate for evaluation.Source: Actor–Critic Learning from Trial History · p. 4

Experimental Setup

• Seven methods share a nine-family conditional search space and aligned evaluation conditions, while candidate generation, priors, initialization, and internal objectives remain method-specific.The shared conditions include a chronological evaluator, attempted-trial budget, forecast-output validity checks, and validation-MASE incumbent rule.Source: Experimental Setup · p. 4

Tasks, Search Space, and Comparators

• The benchmark spans 41 dataset–frequency variants across 13 domains and nine forecasting models, comparing six baselines using random, Bayesian, evolutionary, multi-objective, and language-model-assisted strategies.NSGA-III and MOTPE optimize validation MASE, training time, and parameter count, but incumbents use validation MASE alone; SLLMBO* is a source-based adaptation, not an official reproduction.Source: Tasks, Search Space, and Comparators · p. 4

Evaluation Protocol

• Runs follow each task’s fixed chronological route, charge every attempt to budget, select incumbents by validation MASE, and rank methods within task before averaging across tasks.Test scores neither rank eligible configurations nor break ties, and WQL evaluates the same validation-selected incumbent. Test-MASE trajectories are also summarized by arithmetic mean rank over predefined early, late, and full windows.Source: Evaluation Protocol · pp. 4–5

Full-Budget Forecasting Quality

• TRACE-CASH has the lowest mean rank on both reported full-budget metrics, with WQL assessing the validation-MASE-selected incumbent rather than a WQL-selected optimum.In a separate post-hoc sensitivity using 29 source-group blocks, TRACE-CASH also retains the lowest mean rank on both MASE and WQL.Source: Full-Budget Forecasting Quality · p. 5

Performance Across Trial Budgets

• TRACE-CASH has the lowest mean rank in the predefined late and full windows of the descriptive test-MASE budget profiles.Table 3 summarizes each trajectory by its window-averaged rank; Figure 2 displays mean test-MASE rank over attempted trials.Source: Performance Across Trial Budgets · p. 5, Figure 2 · p. 5, Table 3 · p. 5

Discussion

• The seven-method evaluation assesses TRACE-CASH as an integrated search procedure, including learned candidate generation, exploration, and fixed search rules.The reported results characterize the complete procedure rather than any component in isolation.Source: Discussion · p. 5

Conclusion

• TRACE-CASH has the lowest mean rank on MASE and WQL across 41 task variants.This ordering is retained in a separate post-hoc sensitivity using 29 source-group blocks.Source: Conclusion · p. 5
• TRACE-CASH also has the lowest descriptive window-averaged test-MASE rank in the predefined full and late windows.Source: Conclusion · p. 5
• The complete procedure combines trial-history-conditioned grouped candidate generation with fixed model-coverage, validation-guided exploitation, and recovery rules.Overall, it is competitive among the evaluated TS-CASH methods under aligned evaluation conditions.Source: Conclusion · p. 5