PS is just one big Neural Network.. There's without a doubt a statistically "best" team for each and every archetype

[AquaFox]

For starters I apologize if this is in the wrong thread. I'm not quite sure where to post such a meta topic at.

But whenever I create a team from scratch, the team usually deviates quite a bit from the original approach. Eventually, whenever I get to a point where I feel like I can no longer make any progression on bettering my team in any way, I look up some teams on the same playstyle. For example, I noticed this back in XY when I had an offensive sand team using Excadrill. Eventually the metagame and the team kind of "forced" me to build a team that I considered to be the most "optimal" for such a team. I looked up other teams based around the same idea and sure enough they all looked nearly identical to mine. Then during gen 8 I decided to create a stall team. I had tinkered with it for a while and eventually I ended up with a team that looked identical to every other gen 8 "stall" team. And while some moves/EVs/items here and there were a bit different, overall 95% of the team was identical. This phenomenon reminds me of VGC, probably moreso there, since it's extremely noticeable there. A lot of VGC teams consist of nearly the same pokemon, some were even the same team.

It reminds me a lot of neural networks and how they work. Basically neural networks run thousands/millions of simulations (or in Pokemon Showdown's case, the neural network is just every single user, and the simulations are just every single battle) until eventually an optimal outcome is reached to where no more progress can be made. It really makes me think about the simularities between neural networks and pokemon showdown and whenever I notice how similar my teams to other teams using the same strategy, it makes me really think that there definitely is somewhere out there, a statistically "optimal" team for that archetype and all the battles we have done are just the equivalent to the simulations a neural network would have run.

 

[ironwater]

Hey AquaFox, welcome to Smogon. I’m gonna be honest, I don’t know where something like this would fit. As for this thread I’ll lock it since you are not allowed to start new threads in this subforum without moderator approval and we have existing threads for Questions/Metagame thoughts and discussions, even if they are only about OU and you are talking about something way more general here. As a compensation I’ll give you my thought on this directly in this thread.

I’ll split my answer in two parts, one about teambuilding and one about neural networks to try giving my thoughts on all the things you’ve mentioned.

I think you are viewing teambuilding as something way too static, which, if it was indeed, would mean there’s likely some optimal state to find. However, I don’t think it is the case nor that it is a reasonable approximation for most metagames. Innovation has always been a big part of teambuilding and if we take a look at gen 8 OU for instance, even after the DLCs and without any exterior arrival in the metagame, the tier evolved quite a lot meaning that a team from one year ago may not be as good as it was before. More generally, an issue with seeing teambuilding as something static is that, if we assume there was an optimal team for each archetype, then it will make teams having a favorable matchup against these way better, and there’s always some due to the nature of the game itself. These new teams beating the optimized teams will increase in popularity replacing the old ones and people will start finding ways to beat these new teams. To say it differently, what’s good in a metagame is what win against most of the teams used in this metagame, which explain why new teams emerged as they are good against what is currently used making the pool of good teams used change and thus leading to new metagame and teambuilding developments. You can probably find a team statistical optimal against what is used at a current point in time, but this is deemed to evolve in a complex predator-prey kind of equilibrium.

Now about Neural Network, I wouldn’t say that they run simulations, even if I get what you have in mind there. I would rather say that they see a lot of examples of something with a particular information on them that they’ll try to infer on new examples. Could this be applied to what you mentioned? Well, if we take the “teambuilding is static” hypothesis, which I don’t agree with as explained above, then we can see training examples as teams with how well they did in different games and the goal would be to infer a score on how good the team is. There’s several issue here as for such an algorithm to learn something we would need to have somewhere a non-biased information about the team performance in itself. But if you take replays, there’s also the matchup in the current game (which means you would probably need a ton of replays for every single team) and most importantly there’s also the player as a bad player may lose with a good team against a way worse team. Another issue is that, in a dynamic view of teambuilding, you would need to retrain your model continuously since, as it learns on past data, it will start becoming obsolete as soon as you find new “optimized teams”. Now, I may not have understood what you meant by simulation and if you meant something completely different (eg: make a neural network learn how to play the game and then simulate a huge number of battles, in which case you would not use neural nets to learn building directly, but this is a very hard problem looking at the complexity of the game).

I’m not the most expert person on neural networks, but I’ve already used some and I have a good idea of how they work, and to me I don’t think you can easily use one to find out optimal teams in a metagame, and metagames being dynamic makes this even harder.

This was a bit long and, in the end, just my opinion on this, but I hope it gives some more elements on the topic (and I’m sorry for not being able to tell you where this could fit on Smogon). Anyway, have a nice day!