Note however that tournaments often limit running time (eg 5s for 1000 games on a not-very-fast processor), so you have to be careful with overly complex strategies, like neural nets that are too big.
Until I read this footnote, I was going to suggest throwing the last tournament’s thousand-round matches into a dataset and training a TITANS-based architecture on it until performance against a representative set of bots grabbed from the internet starts to drop from overfitting. Even so, I think it’d be funny to try this out just to see how high you can get the number to go.
With model distillation, you could try squeezing it into something that technically, barely manages to finish its games on time.
Yeah I suspect a distilled/pruned neural-net could do really well against the existing strategies, I dunno what the latest state of the field is. I also don’t have a very good sense of recent progress in computer/ML optimizations so maybe there are other tricks I’m discounting.
Until I read this footnote, I was going to suggest throwing the last tournament’s thousand-round matches into a dataset and training a TITANS-based architecture on it until performance against a representative set of bots grabbed from the internet starts to drop from overfitting. Even so, I think it’d be funny to try this out just to see how high you can get the number to go.
With model distillation, you could try squeezing it into something that technically, barely manages to finish its games on time.
Yeah I suspect a distilled/pruned neural-net could do really well against the existing strategies, I dunno what the latest state of the field is. I also don’t have a very good sense of recent progress in computer/ML optimizations so maybe there are other tricks I’m discounting.