# How to implement the regret matching algorithm?

My question is the following: How to calculate the regret in practice?

I am trying to implement the regret matching algorithm but I do not understand how to do it.

• First, I have $n$ players with the joint action space $\mathcal{A}=\{a_0, a_1,\cdots,a_m\}^n.$
• Then, I fix some period $T$. The action set $A^t\in\mathcal{A}$ is the action set chosen by players at time $t$. After the period $T$ (every player has chosen an action). So I get $u_i(A^t)$.
• Now the regret of player $i$ of not playing action $a_i$ in the past is: (here $A^t\oplus a_i$ denotes the strategy set obtained if player $i$ changed its strategy from $a'_i$ to $a_i$) $$\max\limits_{a_i\in A_i}\left\{\dfrac{1}{T}\sum_{t\leqslant T}\left(u_i(A^t\oplus a_i )-u_i(A^t)\right)\right\}.$$ I do not understand how to calculate this summation. Why there is a max over the action $a_i\in A_i$? Should I calculate the regret of all actions in $A_i$ and calculate the maximum? Also, In Hart's paper, the maximum is $\max\{R, 0\}$. Why is there such a difference?

I mean if the regret was: $\dfrac{1}{T}\sum_{t\leqslant T}\left(u_i(A^t\oplus a_i )-u_i(A^t)\right),$ the calculation would be easy for me.

The regret is defined in the following two papers  (see page 4, equation (2.1c)) and  (see page 3, section I, subsection B).

1. A simple adaptive procedure leading to correlated equilibrium by S. Hart et al (2000)
2. Distributed algorithms for approximating wireless network capacity by Michael Dinitz (2010)

I would like to get some helps from you. Any suggestions step by step how to implement such an algorithm please?