916278a640
Add RL auction agent (Double DQN) and Set Transformer for team valuation
...
- rl_auction_agent.py: Custom Gym-free auction environment with N opponents,
numpy-only Q-network with manual backprop, Double DQN policy with target
network, full training loop with epsilon decay and periodic evaluation,
and baseline comparison vs greedy and MILP strategies.
- set_transformer.py: Team-level valuation model treating roster as an
unordered set. Two modes: full PyTorch Set Transformer with ISAB/PMA
when torch is available, or sklearn Bag-of-Players fallback using
per-role aggregates, pairwise cosine similarities, and position entropy.
2026-08-11 17:27:26 +08:00
3b065775f5
Major refactor: Fantabeto 26/27 — modular package, GBM ensemble, MILP/MCTS optimization
...
Phase 1: Data Engineering
- Refactored notebooks into src/{scraper,features,models,optimization,bot}
- FBref scraper with proxy rotation + Playwright Cloudflare bypass
- Fantacalcio.it integrated scraper (authenticated API + HTML fallback)
- api-football RapidAPI client for supplementary xG/xA/injuries
- RAG news pipeline: Gazzetta, Sky Sport, Di Marzio → injury/suspension/tactical extraction
- 26/27 season config: teams, scoring rules, name mappings, news sources
Phase 2: SOTA ML Architecture
- GBM Ensemble (LightGBM + CatBoost + XGBoost) with stacked blending
- Bootstrap ensemble for uncertainty quantification
- SinhArcsinh distribution head (ported from original TF Probability)
- Card classifiers (yellow/red), penalty model, goal probability (Poisson)
- Temporal GNN for player interaction modeling (crosses→goals, passes→assists)
- Optuna hyperparameter tuning with time-series CV
Phase 3: Operations Research
- Auction solver: MILP knapsack with PuLP (budget + role constraints)
- Grid Auction (Asta a Griglia): Minimax game theory bidding strategy
- Weekly lineup optimizer: MCTS maximizing win probability vs opponent
- Modificatore Difesa integration + captain selection
- Transfer market analyzer: buy-low/sell-high via xG regression to mean
- Opponent behavior modeling from historical lineage patterns
Phase 4: Agentic Workflow
- Telegram bot: auto-briefing (Friday + Sunday morning)
- Tactical briefing generator with start/sit recommendations
- GitHub Actions CI/CD: scheduled pipeline (scrape → predict → notify)
Infrastructure:
- 31 pytest unit tests (features, models, scraper, optimization)
- requirements.txt (lightgbm, catboost, xgboost, optuna, pulp, playwright, langchain)
- Makefile with install/test/lint/scrape/train/bot targets
- Jupyter notebook: 26_27_strategy.ipynb demonstrating auction + matchday 1 mockup
- Completely rewritten README.md with architecture diagram
2026-08-11 13:16:07 +08:00
ramseshk