This repo simulates several online learning algorithms on games. The algorithm in arXiv:2204.11417 is implemented, and the experiment results in the paper is reproduced.

Under general-sum games with <=4 players, each with <=3 actions, gameplays are simulated with different learning rules (EG, OG, vanilla GD, and the two algorithms in the paper).

The utilities are i.i.d. sampled from a uniform distribution. OG and EG both uses orthogonal projection to the probability simplex.

Negative results on OG and EG:

OG and EG both have linear swap regret.
The paper [arXiv:2204.11417v1] claims " $O(\log T)$ second-order path lengths" with their algorithm. OG and EG both have linear path lengths, implying that they don't converge.

Positive results on OG and EG:

I have been messing around with different parameters, and never found a case with a $\Omega(1)$ external regret for either OG or EG, under fixed learning rates. All cases that I've experimented with fixed learning rates have finite external regret. (This does not hold for normal GD.)
Sometimes the swap regret grows linearly, yet the external regret grows negative-linearly in $T$.

Under learning rate $\eta ∝ 1/T^k$ or $\eta ∝ {(1-\varepsilon)}^T$, the external regrets are unbounded.

Reproducing the Paper:

The algorithms OFTRL-LogBar and BM-OFTRL-LogBar in the paper [arXiv:2204.11417] are implemented. The former has O(log T) regret, and the latter O(log T) swap regret.
Surprisingly, it seems that the former also has logarithmic swap regret. Sadly I can't prove that.