meiji163 / bandits

Multi-armed bandit algorithms

Multi-Armed Bandit Algorithms

RNN (LSTM or GRU)

• guesses the opponent policy distribution based on sequence of previous moves
• difficult to learn all possible opponent strategies (deeper network?)

UCB (Upper Confidence Bound)¹

• balance exploitation/exploration
• predictable by opponent
• needs high exploration constant and/or frequent resets

PUCB (Predictor + Upper Confidence Bound)²

• predictor (ideally) detects changes in opponent stategy

SER4 (Successive Elimation Rounds with Randomized Round-Robin and Resets)³

• runs several random trials to find move with highest mean reward
• assumes constant oppponent distribution
• bad against high variance strategies

EXP3.R (EXP3 with Resets)^{3 4}

• updates probabilities based on mean rewards and prior
• reset based on detection of maximum mean reward drift
• good against exploitation-biased strategies

Bayesian (Thompson Sampling)⁵

• use beta distribution to model reward probabilities and update based on observations
• assumes constant opponent distribution

References

1: https://link.springer.com/article/10.1023/A:1013689704352

2: https://link.springer.com/article/10.1007%2Fs10472-011-9258-6

3: https://link.springer.com/article/10.1007/s41060-017-0050-5

4: https://cseweb.ucsd.edu/~yfreund/papers/bandits.pdf

5: https://arxiv.org/pdf/1707.02038.pdf