About AI Battle Replays
The battle simulator reproduces Pokémon battles in a simulator environment, predicting outcomes between various teams through AI-vs-AI battles — designed to discover your team's strengths and weaknesses, track meta trends, and suggest better team compositions.
Target format: Pokémon Champions Singles. Trained battle AIs fight each other and results are analyzed, letting you verify win rates and matchups with "battle data" rather than "gut feeling."
AI Battle Level
The battle AI uses a neural network (policy + value) trained through self-play (AlphaZero-style), combined with Monte Carlo Tree Search (MCTS). It reads promising moves deeply through many simulations and selects the move with the highest predicted win rate. The opponent's build and actions are not hardcoded — they are probabilistically estimated from usage data.
- Opponent build estimated from usage rates: Items, natures, moves, and EVs are probabilistically assumed from usage data, then narrowed down based on damage received during battle (treats battle as imperfect information).
- Strength benchmark: Clearly outperforms simple rule-based AIs (beginner level). The goal is "roughly equal to or slightly above intermediate players" — advanced tactics like bluffing and cycling are still in development.
- Moves are deterministic: Always selects the move it considers best in a given position (in top-tier simulations, battle content varies due to randomness in team selection and builds).
- Strategies not yet mastered: Advanced strategies involving Baton Pass ability inheritance and Illusion (disguise) are still being learned.
※ This is not a perfect optimal solver. Please treat results as "battle data from strong AI matchups" and use them to identify trends.
Top Tier Simulation (Round-Robin)
The target is top-environment players' teams whose builds have been publicly shared. Your selected team is matched against each of those players' teams in a round-robin format, and all battle records are saved. From the results you can check:
- Win rate & favorable/unfavorable matchups: Win rate against each opponent player.
- Team selection: The frequency of which 3 Pokémon were actually selected (yours and the opponent's).
- 1v1 matchup table: Damage dealt/received and win/loss across all 6×6 combinations.
- In-game loss causes: Unkillable cores, frequently knocked-out members, setup sweep tendencies.
- Replays: All saved battles can be reviewed and replayed afterward.
How It Works
| Item | Details |
|---|---|
| Target | Pokémon Champions Singles (Lv50, 32-scale EVs) |
| AI | Policy+Value network + Monte Carlo Tree Search (MCTS) |
| Battles per card | Multiple battles (reflects variation in team selection, builds, and RNG) |
| Draw | Draw if unresolved after 30 turns |
Notes & Limitations
- Win rate granularity: Win rate is shown per number of battles; more battles = finer granularity.
- Team selection perspective: Even if your 6 members have good matchups overall, the selected 3 may vary and affect the outcome. Current analysis includes selection data, but optimizing across all selection patterns is a future goal.
- AI strength: May not match expert-level play involving advanced reads. Use the data as a reference for trends.
Roadmap
Building on the current top-tier simulation (battle evaluation) as a foundation, we will continue adding features and improving existing ones.
Feature Addition: Team Suggestion & Improvement Analysis
- Improvement suggestions: AI analyzes your 6 members and suggests move/item/nature/EV/member changes to improve your win rate.
- Partial team proposal: Fix some members and have the AI suggest the remaining slots based on the current meta.
- Win rate re-evaluation: Re-simulate with the proposed team and compare win rates before and after the changes.
Improvements
- Season adaptation: Update battle AI and target teams for new season usage and build data, and add simulator support for newly available Pokémon and items.