Notes on: Lanctot, M., Lis\`y, Viliam, & Bowling, M. (2014): Search in imperfect information games using online monte carlo counterfactual regret minimization

Table of Contents

1 Notation

  • Chance \(c\) refers to the fixed stochastic strategy which is the environment (i.e. which card comes up on the flop in Texas Hold'em)
  • \(h \in H\) denotes a history
  • \(i \in N\) indexes the players, and \(P(h)\) denotes the player who's turn it is next
  • \(z \in Z \subseteq H\) denotes a terminal history, which is a full game from start to end
  • \(u_i(z)\) denotes the payoff for player \(i\) from the terminal history \(z\)
  • \(a \in \mathcal{A}\) denotes the action space and \(A(I)\) denotes the set of possible actions given infoset \(I \in \mathcal{I}_i\)
  • \(ha\) denotes the history \(h\) followed by action \(a\), and \(h \sqsubset h'\) is notation for "sucessor history"

    \begin{equation*} h' = ha \end{equation*}
  • Probability of terminal history under \(\sigma\)

    \begin{equation*} \pi^{\sigma}(z) = \prod_{i \in N} \pi_i(z) \end{equation*}


    \begin{equation*} \pi_i(z) := \prod_{ha \sqsubset z, P(h) = i} \sigma_i \Big( I(h), a \Big) \end{equation*}

    I.e. the joint probability of the strategies by each player \(i\) in the case where the final action was taken by player \(i\)

  • \(\sigma_{-i}\) and \(\pi_{-i}^\sigma\) refer to player i's opponent strategy and products of the opponent's and chance's actions
  • Behavioral strategy for player \(i\) is

    \begin{equation*} \begin{split} \sigma_i: \quad & \mathcal{I}_i \to \mathcal{P} (\mathcal{A}) \\ & I \mapsto \sigma_i(I) := p \Big( A(I) \Big) \end{split} \end{equation*}
  • Counterfactual value:

    \begin{equation*} v_i(I, \sigma) = \sum_{(h, z) \in Z_I}^{} \pi_{-i}^{\sigma}(h) \ \pi_i^{\sigma}(h, z) \ u_i(z) \end{equation*}


    \begin{equation*} Z_I = \left\{ (h, z) \mid z \in Z,\ h \in I, \ h \sqsubset z \right\} \end{equation*}

2 Stuff

An \(\epsilon \text{-Nash equilibrium}\), \(\sigma\), is a ordered tuple \((\sigma_1, \dots, \sigma_N)\) such that for any given player \(i\), the benefit of switching strategy from \(\sigma_i\) to some \(\sigma_i'\) is bounded by \(\epsilon\), i.e.

\begin{equation*} \Big( \max_{\textcolor{red}{\sigma_i'} \in \Sigma_i} \underset{\sigma_1, \dots, \textcolor{red}{\sigma_i'}, \dots, \sigma_N}{\mathbb{E}} [u_i(z)] \Big) - \underset{\sigma_1, \dots, \textcolor{green}{\sigma_i} , \dots, \sigma_N}{\mathbb{E}} [u_i(z)] \le \epsilon \end{equation*}

If \(\epsilon = 0\), it's called a Nash equilibrium.

A zero-sum game is a game such that

\begin{equation*} \sum_{i=1}^{N} u_i(z) = k, \quad \forall z \in Z \end{equation*}

for some constant \(k\).

The exploitability of a profile \((\sigma_1, \dots, \sigma_N)\) is defined

\begin{equation*} \epsilon_{\sigma} = \sum_{i=1}^{N} \max_{\sigma_i' \in \Sigma_i} \Big( \underset{\sigma_1, \dots, \sigma_i' , \dots, \sigma_N}{\mathbb{E}} [u_i(z)] \Big) \end{equation*}

i.e. difference between the optimial strategy for each player given the strategies of the other players.