I once ended up participating in rock, paper scissors strategy competition on a whim.

At that time I was quite tired of polishing and repolishing a paper draft at my university. (I write very poorly). And there it was, an announcement of this fun competition. The deadline was just an hour away.

I had no time to implement any respectable algorithm myself.

So, all my submission did was take the history of plays of my opponent, up to the present point in the ongoing joust, and extend it by three possible completions. I would compress each of the three resulting strings. Whichever completion gave the shortest compressed final string, I assumed that to be my opponent's predicted play. I played whatever beat that predicted play.

This did not win the competition, but survived the knockout for many rounds, beyond my expectation. If I remember correctly, I had used plain old zip. A compressor that converges faster would have been better.

In essence my strategy was similar to a Markov chain predictor. It would have been exact match had I used a Markov model based compressor.

The number of rounds of play against one's opponent was not long. Had it been so, Lempel-Ziv being an universal compressor, would have been hard to beat.

Of course we know from game theoretic analysis what the Nash strategy is for this game, but where is the fun in playing that ?

One might think that a transformer based player would be an excellent strategy, not necessarily. One needs to learn potentially non-stationary strategies of the opponent from very few examples - (near) zero-shot online learning is required.

If I had more time, I would have tried context-weighted trees - universal learners for stochastic strings. The failure mode would be non-ergodic plays by the opponent, assuming that, it too is another stochastic parrot.