Hacker News Reader

Comment by nbkvjones | Hacker News Reader

there was a "discussion" on nadota.com about the bot and the semipro player that openai used to test chimed in.

apparently the set of items the bot chose to purchase from was limited[1] and recommended by the semipro tester. As someone who knows next to nothing about ai, my question is this: the bot was announced on stage as blank slate, dumped into dota, and built entirely from grinding countless games against itself; is it reasonable to pitch it this way while having this item constraint from an outside source? I also wonder what else was recommended by the tester, and then constrained.

the "discussion" is linked below and the tester is the user sammyboy. Here's a warning though: nadota is 99% trolling, hate, idiocy, and garbage.

[1] http://nadota.com/showthread.php?41718-terrifying-1v1-mid-AI...

View on HN

(I work at OpenAI.)

We'll have another blog post coming in the next few days. But as a sneak peek: we use self-play to learn everything that depends on an interaction with the opponent. Didn't need to with those that don't (e.g. fixed item builds, separately learned creep block).

Can you give us a sneak peak of OpenAI's approach to 5v5 Dota? :D

So it's not exclusively unsupervised self-play? This contradicts how it was portrayed during the game/afterwards.

See the opening video (https://www.youtube.com/watch?v=wiOopO9jTZw&t=268), where I mention that we "coach" it on what we think is good and bad. All the heavy lifting is self-play (which I would characterize as supervised).

That's good to hear.

There's probably a huge disconnect between what an AI engineer entails with a certain word and what the layman then comprehends. At the same time, I think we (nadota) probably latched onto certain words and downplayed the ones that would make it more impressive to us. Just a result of itching for information without there being much out there.

Really cool work, really glad you're doing it in dota, and eagerly awaiting your blog post.

https://blog.openai.com/more-on-dota-2 A new blog is already here.

The other big "hardcoded" constraint is most likely the usage of the bot API. The API itself is complex with lots of functionality and I would assume that researchers were extremely smart about picking out just the right API calls that are needed to make it work. That's very different from and much easier than training an agent based on raw keyboard/mouse input.

> is it reasonable

you ask the wrong question. Sure it's reasonable, noone's going to sue them

Reply on news.ycombinator.com