Case study / 06
Learn to rank the move, not merely imitate it.
A simulation and ML pipeline evaluating legal moves through reproducible game experiments.
- Role
- ML Engineer
- Scope
- Simulation · Feature design · Training · Evaluation
- Status
- Research prototype
- Year
- 2026
01 / CONTEXT
The problem
Generate useful simulated data and evaluate decisions when several legal moves may be reasonable.
02 / RESPONSE
The response
A JSONL pipeline connects seeded simulation, vectorization, supervised models, candidate-value scoring, and matched-seed gameplay.
03 / EXPERIENCE
Product flow
- 01
Simulate
Generate reproducible trajectories.
- 02
Vectorize
Encode state, hand, and legal actions.
- 03
Train
Compare three modeling strategies.
- 04
Rank
Score each legal candidate.
- 05
Evaluate
Measure accuracy, regret, and win rate.
05 / SYSTEM
System architecture
Simulation, learning, and gameplay evaluation remain replaceable and reproducible.
06 / TRADE-OFFS
Architecture decisions
Candidate value
Ranking legal moves matches the real decision.
Matched seeds
Bots face comparable randomness.
Multiple metrics
Accuracy is supplemented by regret and gameplay.
07 / QUALITY
Security & reliability
- Seeded games make comparisons repeatable.
- Legal-action filtering prevents impossible predictions.
- Serialized models preserve experiments.
08 / EVIDENCE
Outcomes & evidence
- A complete path from simulation to playable bot.
- Comparable baselines for three strategies.
- Evaluation grounded in prediction and gameplay.
09 / STACK