Comment by sillysaurusx | Hacker News Reader

Comment by sillysaurusx | Hacker News Reader

Parent Full thread

Lmao, imagine calling AdmiralBulldog’s primary strategy “cowardly”.

Sorry for the rare persona break, but Dota happens to be my old passion. It was a bit like saying that Magnus wouldn’t play a certain chess strategy because it would be “cowardly.” I assure you, the top teams are there to win. They also saw how ungodly effective OpenAI’s 1v1 bot was, and everyone respected its skill.

You can make those claims, but I think we’ll have to agree to disagree. More specifically, I agree that OpenAI solved a limited subset of Dota. But I disagree that OG would have “figured it out” under the specific rules the bot was trained on. It’s a totally different game, and it’s nothing like bottle crow (which I was fond of back in 2010). It also wasn’t bottles, because only the mid lane buys a bottle. The other lanes were usually health regen to prevent your cores from having to go back to base - which of course the AI did here, but to a ridiculous degree that no one was expecting.

Which is more likely: that Dota happens to be the one game that AI can’t become superhuman at, or OpenAI didn’t train quite long enough to prove it?

Yep, we'll have to agree to disagree.

One reason I disagree is that Dota (and Starcraft 2) are both games of imperfect information (more like Poker than Chess or Go).

That means a lot of the time, the game revolves around deception and subterfuge. If you've ever had to hunt down an enemy splitpusher and have to decide between spending 10s checking a hiding place and having a chance at getting a potentially game-winning kill, but also having a chance of not finding them and wasting your time, you'll know what I'm talking about. There's no optimal strategy, because all strategies have weaknesses to be exploited (e.g. In BW, if Terran knows you as Protoss always go greedy such as 12 Nexus, they can punish you with BBS. However if you play safe, they can play greedy themselves with 14CC, so there are a lot of mindgames.)

After AlphaZero's victories or DeepBlue's for that matter, professional Go and Chess players could find no weaknesses. Nada.

After OpenAi 5 became available, a professional player compiled a list of 20 weaknesses that I'm not certain OpenAI can ever fix (see link above). How do you tell the neural net when to dust or ward?

The same story with DeepMind's AlphaStar. Despite playing a good macro game, players online have figured out lots of ways to cheese it (e.g. send just one unit will cause all workers to be pulled and stop mining). I understand Poker's been "solved", but these computer games with much larger action spaces and imperfect information might turn out to be significantly harder to pin down than Chess or Go. There are enough edgecases that it's a much better representation of real life problems like self-driving cars (negotiating a merge for example).

>How do you tell the neural net when to dust or ward?

My knowledge on what is feasible with AI is fairly limited, so maybe a proposition like this is a bit naive - but wouldn't a more effective solution be to use some sort of machine learning/neural net hybrid that also incorporates data from matches played high-skill teams?

From what I'm reading, OpenAI trains the neural net entirely by playing games against itself, and only uses matches against humans as benchmarks - so it essentially only develops strategies of play that primarily work well against itself. In most cases those probably coincide with strategies that work against human teams, but it seems like a lot of information is going to go missing - which is probably why it's only been successful in certain subsets of the game.

It reminds me of a story I read a while back about two children who grew up relatively isolated from the rest of society, but had access to musical instruments. Since they had no teacher and apparently no other training materials, they developed an entirely unique style of performance and composition. Obviously in the case of music "success" is entirely subjective so simply being different doesn't invalidate it. But the point of OpenAI is to eventually beat human players in all aspects of the game - and if it's not going to actually train against humans, then success is going to be at least partially coincidental. So I would definitely agree that OpenAI (as it stands right now) has some potentially insurmountable weaknesses.

I don't think it changes the fact that given enough compute time, there is a "GTO" strategy for imperfect information games like Dota, just as there is for poker. The AI will lose some proportion of games, but overall it'll win.

https://www.deepstack.ai/

Reply on news.ycombinator.com