Dota 2 Fantasy Analytics & Decision System

Python · PyTorch · Transformers · Reinforcement Learning · Monte Carlo Planning · SQLite · Streamlit

An end-to-end research software project for tournament fantasy analytics and sequential Dota Fantasy banner optimization. It combines reproducible esports data pipelines with a rule-faithful simulator, Transformer actor-critic, and actor-guided Monte Carlo planning.

The project separates real tournament data from synthetic decision trajectories. Training metrics never promote a model by themselves: every candidate requires an independent paired evaluation on matched seeds.

At a glance

Area	Current scope
Tournament analytics	157 EWC 2026 maps, player-map fantasy statistics, role aggregation, provenance and coverage tracking
Sequential optimizer	Exact 30-step simulator, three token offers plus refresh, masked variable action space
Active decision system	Validated baseline Transformer actor plus bounded Monte Carlo planner
Promotion protocol	Matched schedules, paired deltas, bootstrap confidence intervals, safe/balanced/ceiling scenarios

Key results

Tournament analytics

• 157 EWC 2026 maps are available in the local compact database.
• The data layer supports player-map, role-slot, stage, source-coverage, and per-stat fantasy analysis.
• The fact agent uses a source-first SQL route for database-backed facts.

RNG optimization

• The active system is baseline actor + exact Monte Carlo planner.
• A conservative actor-critic candidate was rejected after an independent safe-scenario regression, despite acceptable training metrics.
• Historical generated reports were intentionally pruned. Do not publish stale candidate deltas: generate a fresh matched result via docs/RESULTS.md.

Product surfaces

The Streamlit Fantasy Hub exposes the research system through four focused workflows rather than one monolithic screen.

Human vs Model	Live Advisor
Shared-offer 30-step simulation against the planner.	Screenshot OCR, manual correction, token entry, and a guided move.

Banner Analysis	Stats & Players
Per-emblem multiplier and role-stat benchmark diagnostics.	Historical fantasy maps, ceiling rankings, reliability, and role pairs.

System architecture

flowchart LR
    A[Esports sources] --> B[SQLite + provenance]
    B --> C[Analytics views and fantasy scoring]
    C --> D[Role/stat benchmarks]
    D --> E[Exact 30-step RNG environment]
    P[Empirical starter and token presets] --> E
    E --> F[Teacher trajectories]
    F --> G[Transformer actor / Q / V]
    G --> H[Actor-guided MC planner]
    H --> I[Matched-seed promotion gate]
    I --> J[Streamlit advisor]

Loading

See Research architecture for tensor representation, PPO, offline RL, planner design, and evaluation methodology.

Quickstart: portable demo

The demo needs no tournament database, model artifact, API token, or OCR.

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -U pip
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe scripts\run_demo.py
.\.venv\Scripts\python.exe -m streamlit run examples\demo_app.py

scripts/build_demo_db.py creates an ignored deterministic SQLite database at data/sample/demo.sqlite. The demo is synthetic and exists solely to make a clean clone reviewable.

Full local setup

The full advisor needs ignored local artifacts:

data/ewc_2026_fantasy_compact.sqlite
data/ti_2026_fantasy_compact.sqlite
models/rng_neural_slot_selfplay_selected_v1.pt

Then run run_rng_human_vs_model_ui.cmd. For screenshot OCR, run run_rng_ui_ocr_setup.cmd and install system Tesseract.

Reproducible evaluation

.\.venv\Scripts\python.exe scripts\run_portfolio_evaluation.py `
  --candidate models\candidate.pt --episodes 100

The generated report is ignored by Git. Record only candidates that pass the promotion gate in docs/RESULTS.md.

Repository layout

app/        Streamlit advisor and dashboards
configs/    tournament metadata, RNG presets, title rules
data/       ignored local artifacts plus portable demo builder
docs/       methodology, architecture, results protocol
examples/   self-contained demo
notebooks/  collection and fact-agent workflows
scripts/    supported collection, validation, and training CLI
src/        data, scoring, simulator, planner, and ML implementation
tests/      deterministic rule and smoke tests

Documentation

• Data and SQLite
• Analytics and fact-agent
• Research architecture
• RNG actor-critic and planner
• Results and evaluation protocol
• Reproducibility

Research questions

1. Which player and role combinations offer robust fantasy upside rather than a single-map outlier?
2. Can a rule-faithful simulator improve sequential banner decisions under stochastic token offers?
3. Does an actor-guided planner outperform its baseline under matched random schedules?

Training loss is diagnostic. Independent paired evaluation is the decision criterion.