Agent Evolution
Baseline has two main variants:
- Blueprint: offline-solved strategy library (no live solving).
- Baseline AI: live depth-limited solving on top of a biased blueprint, then reach subgame solving.
Benchmark setup
- Game: Heads-Up No-Limit Hold’em
- Stacks: 200 big blinds, no rake
- Opponent: Slumbot (API version noted per run)
- Metric: BB/100 = big blinds won per 100 hands
Results against Slumbot
| Agent |
Year |
Hands |
Result (BB/100) |
| DeepStack (DeepMind) reimplementation |
2017 |
150,000 |
-6.3 |
| ReBeL (Meta) |
2020 |
- |
+4.5 |
| Blueprint v1 |
2023 |
100,000 |
+3.4 |
| GTO Wizard |
2023 |
150,000 |
+19.4 |
| Blueprint v2 |
2024 |
20,000 |
+10.9 |
| Baseline AI |
2024 |
10,000 |
+18.1 |
Changelog
Blueprint v1 → v2
- Potential-Aware Cards Abstraction
- Smarter betting tree, using Pseudo-Harmonic Mapping not only to translate actions but also to make the betting tree itself less exploitable.
- (Proprietary) Improved betting tree, based on an analysis of blueprint, instead of theoretical harmonic thresholds.
- (Proprietary) Smarter action-translation, also based on an analysis of blueprint.
Blueprint v2 → Baseline AI
- Depth-limited solving: Solves the subgame in real-time instead of only following a stored policy.
- Biased blueprint at the depth limit: The blueprint acts as a prior / continuation model so live-solving doesn’t hinge on a single assumed continuation.
- Reach solving: Avoids overfitting to a fixed opponent range by pricing “range switching” (playing irrationally with other hands), keeping the strategy non-exploitable.
- Results: In a short stack setup (HU 25 BB stack), Baseline AI defeats current blueprint by approximately 7.7 BB/100, which corresponds to the difference we observed in Slumbot matches.
