HyperOp

Hyperparameter optimization for genetic programming symbolic regression on Feynman benchmarks and Steel Industry energy data.

Requirements

ResultDTO snapshots. No R bindings or R-specific .NET dependencies are required.

• .NET 10 SDK.
• PowerShell 7 for batch runs.
• Docker Desktop with Linux containers for irace (the image builds automatically).
• Ollama with the requested model available for agent runs. and press F5. This profile runs the .NET experiment on the host.

Run

defaults directly in code or override them with command-line arguments. From the repository root:

dotnet run --project HyperOp.Experiments -c Release -- --algorithm RandomSearch --instance I.6.20 --configurations 10 --evaluations 1000 --output ./results/demo

variables or manual Docker build/start commands are required. The initial build Methods: RandomSearch, GridSearch, HillClimbingSearch, IRaceSearch, AgentSearch. Use --help for all options and benchmark names. The pipeline automatically starts an R container for each optimization run and removes --configurations caps GP trials, including repeats. --evaluations caps candidate evaluations per trial, including initialization; the minimum is 100. Supply both when overriding budgets. irace requires at least four trials. The image runs Rscript as its entrypoint. Docker must be running and able to bind For a local agent: container, with standard input/output carrying proposals and costs. The experiment itself stays on the host. The image is reused across calls; each Execute call dotnet run --project HyperOp.Experiments -c Release -- --algorithm AgentSearch --instance I.6.20 --model qwen3:8b --ollama-url http://localhost:11434 --output ./results/agent-demo

After changing the Dockerfile, rebuild the image manually from the repository root: Agent runs request Continue/Stop decisions every 100 offspring generations. Change DecisionIntervalGenerations and DecisionTimeout in AgentParameter.cs. Short runs may finish before a checkpoint. Controller failures fall back to Continue.

Batch runs

docker build -f docker/irace.Dockerfile -t hyperop-irace .
./scripts/Run-Experiments.ps1 -Algorithms RandomSearch,IRaceSearch -Instances I.6.20,SteelIndustry -Repetitions 10 -Output ./results/campaign01

Run a short experiment:
The launcher builds once, runs CPU jobs in parallel, then agent jobs sequentially,
and combines completed results into `all_results.json`. Use `-MaxProcesses` to set
CPU workers, `-PlanOnly` to preview jobs, and `-RepetitionStart` to assign distinct
repetition ranges across machines. Agent batches accept `-Models` and `-OllamaUrl`.

## Results

Each completed experiment produces a `*.result.json` file; batch runs also save
stdout/stderr logs. Completed jobs are reused. Interrupted experiments restart.
Use a new output directory after changing code, data, prompts or settings.

Tuning uses validation NMSE (lower is better). Test scores are saved for the selected
expression and for the nominated configuration rerun with five fresh GP seeds.
These assessment runs are additional to the tuning budget. Agent results also
record checkpoint decisions and inference times.

Rebuild the aggregate manually:

```powershell
dotnet run --project HyperOp.Experiments -c Release -- --aggregate --output ./results/campaign01

Checks and references

five budget variations, and --algorithm all to run every optimizer. The default is

dotnet run --project HyperOp.Checks -c Release

The search space comes from AlgorithmParameter range attributes and MutatorType. The checks do not need Docker or a model service. See the Feynman specification, Steel dataset notes and paper sources for research details.

For irace, numConfigurations / --configurations means the maximum total target evaluations, including repeated trials of a configuration. A minimum of four is required. Returned results include every actual trial, and ConfigurationIndex stores irace's configuration ID, so repeated IDs are expected. irace can stop before the budget is exhausted. Small budgets use fewer racing iterations (one for ten trials); larger budgets use up to five iterations. GA stopping happens at the library's candidate-evaluation checkpoints; EvaluationsUsed uses the same generation-count approximation as the existing searches.

Adjust --configurations to change irace's total trial budget and --evaluations to change the GA work per trial. Both options must be supplied together. Results are written under results/<run-id>/, including per-experiment JSON and all_results.json.