alirezaaminzadeh/solverport-selection-model

SolverPort Selection Model

Multi-head meta-model for optimization solver portfolio selection — predicts best solver, configuration, runtime, optimality gap, and feasibility probability from 18 instance features.

Method

Ensemble calibrated scoring (LightGBM / CatBoost / Random Forest / Learning-to-Rank style) with expert-calibrated weights. No training required at inference time.

Heads

Head	Output
Solver rank	Top-1 solver recommendation + full ranking
Runtime	Predicted solve time (seconds)
Gap	Predicted optimality gap (%)
Feasible	Probability of finding feasible solution
Config	Auto-tuned solver parameters

Features (18)

General (7) + Graph (5) + Problem-specific (6) — see instance feature dataset.

Solvers

cp_sat, highs, cbc, scip, glpk, gurobi, minizinc, alns

Tunable Parameters

time_limit_sec, num_search_workers, presolve, cut_strategy, search_branching, heuristic_frequency, neighborhood_size, initial_solution

Usage

from huggingface_hub import hf_hub_download
import json

weights = json.load(open(hf_hub_download(
    "alirezaaminzadeh/solverport-selection-model",
    "selection_weights.json",
)))
config = json.load(open(hf_hub_download(
    "alirezaaminzadeh/solverport-selection-model",
    "config.json",
)))

Related

• Space: automated-optimization-solver-portfolio
• Features: solverport-instance-features
• Performance: solverport-solver-performance

License

Apache 2.0