# 2022.5.25 A small CFR experiment

## Settings
### Game
2-player Dudo (each player has 1 die)

### CFR algorithm
Outcome Sampling Monte Carlo CFR (Lanctot et al., 2009)
- Strategies on a sampled path is updated for each iteration

## AI agents
The following $8$ agents are prepared.

### Naive Dudo AI
|#iterations|exploitability|
|:-:|:-:|
|$10^3$|$1.6147998$|
|$10^4$|$1.2615279$|
|$10^5$|$0.7056980$|
|$10^6$|$0.2901173$|

### Abstract Dudo AI
- Abstraction: players forget all player's decisions, except the last one
- Exploitability is computed in an abstract Dudo game.

|#iterations|exploitability|
|:-:|:-:|
|$10^3$|$1.4553667$|
|$10^4$|$0.9087963$|
|$10^5$|$0.3376476$|
|$10^6$|$0.1703836$|

## Result
### Win rate table
- Win rate of the column AI
- $10^5$ trials for each pair

||Na3|Na4|Na5|Na6|Ab3|Ab4|Ab5|Ab6|
|-|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
|**Na3**|-|0.64355|0.70853|0.72941|0.54092|0.60223|0.60282|0.60350|
|**Na4**|-|-|0.59068|0.62911|0.44060|0.50413|0.53785|0.53930|
|**Na5**|-|-|-|0.54494|0.39779|0.46675|0.50873|0.51763|
|**Na6**|-|-|-|-|0.37375|0.45145|0.49207|0.49998|
|**Ab3**|-|-|-|-|-|0.67113|0.69791|0.71040|
|**Ab4**|-|-|-|-|-|-|0.56866|0.58357|
|**Ab5**|-|-|-|-|-|-|-|0.51355|
|**Ab6**|-|-|-|-|-|-|-|-|