Battle Bot Nessie123: An OTS VGC bot that topped the Reg M-C Bo3 ladder

[Dusk Mage Necrozma]

Is it cheating to memorise lines given by an ai, then play them out on ladder?

Click to collapse...

The optimal plays in Pokemon are a spread of probabilities each turn, like how the optimal play in Rock-Paper-Scissors is a 33% probability of playing each hand. Bots here work by (I'm pretty sure) computing those probabilities and then making a weighted die roll using them.

If you have a bot just give you the line of highest-probability moves against a given team, memorize that line, and try to play it out on the ladder, you will be playing worse than the bot, plus your line will of course disappear as soon as your opponent makes a decision that deviates from it.

In certain endgame situations where there does exist a play that always wins, I don't see how finding that ahead of time with a bot is cheating any more than "a top player taught me ahead of time that X play always wins this endgame" is ghosting.

 

[Monai]

it's not cheating it's just fuckin loser shit

 

[Monai]

Are virtually all chess players losers for using chess engines to train? We can be rightly concerned about the imminent problem of bot-assisted cheating without demonizing every aspect of this technology

Click to collapse...

Comparing pokemon to chess is a false equivalency and even bringing up the comparison lends the argument more credence than it deserves. The highest levels of chess are played in person and memorizing complicated lines of play works because the game starts from the same point every time. Even in SPL where me and luckoverskill were playing through lots of openings, I would never copy previous experiences because of how drastically things can change within the first couple of turns for any number of reasons. Even developing this technology for Pokemon is an existential threat to how we play it competitively, but there's nothing we can do about it.

 

[niminm]

How long did the harness take to build

I think the biggest bottle neck stopping millions of bots is building the harness
Did the harness accelerate training time by being gpu based?

 

[sheepie]

one of the biggest arguments i see about the positives of bots like these is that they will help with prep which is like. yeah? i guess. i mean it could be helpful for testing a team over and over until you master mus but like at the same time the amount of interest i have in playing a game where the core of my prep is built around a computer telling me whats optimal in match ups or playing a computer over and over and over is genuinely just 0. like i dont fucking want to do that. i know some people are interested in that and sure you do you but its not a global thing. its not like we all are itching for an ai prep tool but just hate the negative of it being able to single handedly ruin the tour environment (yes im dooming but i dont care and its still entirely possible). its like no thats not at all interesting to me and im sure im not alone, not to mention the sheer quantity of people who will either just become entirely reliant on ai or realise they can just skip the whole prep process and just cheat and its like yeah theres a lot of people that wont do this and will act in good faith but it does take just one bad actor to ruin a whole tour. anw im getting off track but like sure maybe it can be a neat tool for prep for some people but that in no way outweighs the downsides and, to some, or me at least, isnt even an upside to begin with

 

[Kau]

The optimal plays in Pokemon are a spread of probabilities each turn, like how the optimal play in Rock-Paper-Scissors is a 33% probability of playing each hand. Bots here work by (I'm pretty sure) computing those probabilities and then making a weighted die roll using them.

If you have a bot just give you the line of highest-probability moves against a given team, memorize that line, and try to play it out on the ladder, you will be playing worse than the bot, plus your line will of course disappear as soon as your opponent makes a decision that deviates from it.

In certain endgame situations where there does exist a play that always wins, I don't see how finding that ahead of time with a bot is cheating any more than "a top player taught me ahead of time that X play always wins this endgame" is ghosting.

Click to collapse...

That doesn't change what I'm trying to say, and you've only taken out a snippet of my whole point there. I'm saying how can you differentiate between someone actually cheating on the ladder and someone who's just memorised the lines.

And on your point about "your line will of course disappear" once they deviate, that's the whole point of studying the lines. People can memorise many branches of the most relevant lines and already understand and know how the battle can play out before it's even started.

I really don't understand why you've chosen to focus that specific sentence. I'm not asking "Is it cheating?" I'm asking, how are you going to differentiate between a person who's cheating and not, when they play identically thanks to memorising lines beforehand.

 

[HibikiVGC]

How long did the harness take to build

I think the biggest bottle neck stopping millions of bots is building the harness
Did the harness accelerate training time by being gpu based?

Click to collapse...

A harness takes around 2 days to build if you scope it to a specific format - that's definitely not the bottle neck, especially when basically every bot dev already explain in some detail how they built their harness.
The only 'bottle neck' so to speak is $$$ for your AI sub (I doubt anyone is actually building the harness by hand when it's the releast expression of the ''copy Showdown but in Rust, make no mistakes" meme), and owning a decent enough GPU or $$$ for renting one.

You can build a top tier model on a budget of 30$ to 200$ depending on how fast you want it, and if you already own a GPU that has reasonable throughput for training.

 

[niminm]

A harness takes around 2 days to build if you scope it to a specific format - that's definitely not the bottle neck, especially when basically every bot dev already explain in some detail how they built their harness.
The only 'bottle neck' so to speak is $$$ for your AI sub (I doubt anyone is actually building the harness by hand when it's the releast expression of the ''copy Showdown but in Rust, make no mistakes" meme), and owning a decent enough GPU or $$$ for renting one.

You can build a top tier model on a budget of 30$ to 200$ depending on how fast you want it, and if you already own a GPU that has reasonable throughput for training.

Click to collapse...

I was mistaken with this being https://github.com/vjeux/pokemon-showdown-rs
you instantly have an engine with about 500x speeds of showdown
so an engine is not rlly a huge bottle neck

 

[BotMakerGuy]

I only glanced at this thread before but having read it fully now, you have a lot of really interesting ideas, and it seems like they led to a lot of success! 1.5M parameters is tiny compared to mikumiku(8.7M) and mine(7.5M). It makes me wonder if the strength of a bot really comes from more efficient search algorithms. I love how it basically seems to solve endgame situations too. Will definitely try to implement some of this stuff into mine. I really cannot get over how clever the entire search system is to find an equilibrium on each turn and then play from it. Great stuff.

Wait, duh, the smaller model is actually better to get the full gain from a clever search system since you can evaluate positions much faster. Very clever.