A comprehensive tool for assessing AI agents performance in simulated poker environments. Written in Typescript.
Each LLM is benchmared over 1000 hands of No Limit Texas Holdem ($1/$2) $300 Cash Game vs 2 vanilla gpt-4o models. We will contiously be releasing benchmarks for new models/agents, feel free to do PRs with your own benchmarks.
| Rank | Agent | BB/100 |
|---|---|---|
| 1 | mistral-large-latest | +11.26 |
| 2 | gpt-4o | -14.78 |
| 3 | claude-3-5-sonnet-latest | -19.95 |
| 4 | gpt-4o-mini | -45.09 |
| 5 | gemini-1.5-pro-latest | -166.85 |
npm i @superagent-ai/poker-eval
// index.ts
import { PokerGame } from "@superagent-ai/poker-eval";
import { Player, PlayerAction } from "@superagent-ai/poker-eval/dist/types"
// See example agent: https://github.com/superagent-ai/poker-eval/blob/main/examples/ai-sdk/agent.ts
import { generateAction } from "./agent";
async function executeGameSimulation(numHands: number): Promise<void> {
// Setup AI players
const players: Player[] = [
{
name: "GPT 1",
action: async (state): Promise<PlayerAction> => {
// Use any model, framework or code to generate a response
const action = await generateAction(state);
return action;
},
},
{
name: "GPT 2",
action: async (state): Promise<PlayerAction> => {
// Use any model, framework or code to generate a response
const action = await generateAction(state);
return action;
},
},
];
// Setup a game
const game = new PokerGame(players, {
defaultChipSize: 1000,
smallBlind: 1,
bigBlind: 2,
});
// Set the output director for stats collection
const results = await game.runSimulation(numHands, { outputPath: "./stats" });
console.log(`Simulation completed for ${numHands} hands.`);
console.log("Results:", results);
}
// Execute the function with ts-node index.ts
executeGameSimulation(5).catch(console.error);After the hands are completed you can find the the dataset in the outputPath you specified above.
| position | hole_cards | community_cards | bb_profit |
|---|---|---|---|
| UTG | Ah Kh | 2d 7c 9h 3s 5d | 3.5 |
| CO | Qs Qd | 2d 7c 9h 3s 5d | -1.0 |
| BTN | 9c 9s | 2d 7c 9h 3s 5d | 2.0 |
| SB | 7h 8h | 2d 7c 9h 3s 5d | -0.5 |
| BB | 5c 6c | 2d 7c 9h 3s 5d | 1.0 |
In this example, the dataset shows the position of the player, their hole cards, the community cards, and the big blind profit (bb_profit) for each hand. The positions are labeled according to standard poker terminology (e.g., UTG for Under the Gun, CO for Cutoff, BTN for Button). The hole cards and community cards are represented in a standard card notation format, and the bb_profit indicates the profit or loss in terms of big blinds for the player in that hand.
BB/100, or Big Blinds per 100 hands, is a common metric used in poker to measure a player's win rate. It represents the average number of big blinds a player wins or loses over 100 hands. To calculate BB/100, use the formula:
BB/100 = (Total bb_profit / Number of hands) * 100
This formula provides a standardized measure of performance, allowing for comparison across different sessions or players by normalizing the win rate to a per-100-hands basis.
Poker combines elements of strategy, psychology, risk assessment, and partial information - perfect for testing an Agent's decision-making skills in complex, uncertain environments. Poker provides measurable KPIs like EV, BB/100, All-In adj BB/100 and VPIP. These KPIs are widely recognized standards, not created by a single company, making them ideal for objectively evaluating an Agent's decision-making skills.
We've specificalyy chosen No Limit Texas Holdem cash games and are officially calling the eval NLTH.
We've created some examples using populat agent frameworks you can use as inspiration (feel free to contribute):
{
"cff-version": "1.2.0",
"message": "If you use this software, please cite it as below.",
"authors": [
{
"family-names": "Ismail",
"given-names": "Pelaseyed"
}
],
"title": "Superagent PokerEval",
"date-released": "2024-11-25",
"url": "https://github.com/superagent-ai/poker-eval"
}