MIT Poker Bot 2026 - Toss or Hold'em

Team: Anthony Le, Shangun Bothra

Overview

This repository contains the progression of poker bots developed for MIT's Poker Bot competition (2026), competing in a variant called Toss or Hold'em. The strongest iteration is week3_bot, which implements advanced opponent modeling and position-aware decision making.

Game Variant

Toss or Hold'em is a simplified poker format where:

• Players start with 3 cards (preflop)
• Each player discards 1 card (toss phase)
• Two community cards are revealed
• Players play the final 5-card hand

Strategy Overview

Discard Phase

• Monte Carlo simulation to evaluate win rates for each discard option
• Position exploitation: In-position players leverage opponent discard information to refine future decisions
• Entropy modeling: Dynamically adjusts discard strategy based on board uncertainty and opponent tendencies
• Adaptive budget allocation: Simulation count scales with decision importance and time pressure

Betting Strategy

Preflop (3-card):

• Premium hands (pairs 9+, Ace-Broadway): Aggressive 3-betting
• Strong holdings (pairs 7-8, high broadway): Position-dependent raising
• Medium hands: Defensive calling
• Weak hands: Fold to aggression or check/call

Postflop (2-card):

• Strong hands (≥0.75): Value betting with EV-optimized sizing
• Medium hands (0.40-0.75): Semi-bluffing in position, pot-odds adjusted calls
• Weak hands: Positional bluffing with controlled frequency
• Bet sizing: Non-deterministic ±15% variance to obscure hand strength

Opponent Modeling

• Tracks opponent discard patterns (high cards, entropy-adding cards)
• Laplace-smoothed statistics prevent early-game bias
• Adjusts aggression based on inferred opponent tendencies

Results

week3_bot demonstrates consistent performance against week2_bot:

• Mean winnings: ~37,374 chips per game
• Cumulative growth over 100 games shows sustained profitability
• Superior to earlier iterations through exploit and adaptive strategies