<< Back to General Forum   Search

Posts 1 - 8 of 8   
OpenAI on the Multi Day Ladder?!: 9/9/2017 01:04:45

player12345
Level 61
Report
AI/computers now dominate in Chess, Go and recently DOTA2 1v1. Are there any 1v1 strategy games remaining where human champions can still win?

Maybe we still have a chance in multi-template Warlight competition. Do you think an OpenAI powered bot could earn a gold trophy on the MD ladder? http://md-ladder.cloudapp.net/
OpenAI on the Multi Day Ladder?!: 9/9/2017 01:35:06

Fizzer 
Level 58

Warzone Creator
Report
In theory? Yes, it could. In practice? No, we'll never see it in our lifetime, imo.
OpenAI on the Multi Day Ladder?!: 9/9/2017 02:02:43


TeamGuns 
Level 59
Report
^I think you underestimate way too much AI evolution Fizzer. If you had some great programmers with ressources, I would give five years tops until an AI is able to beat human players most of the time. Without those ressources, I'd say ten years.
OpenAI on the Multi Day Ladder?!: 9/9/2017 02:05:05


Njord 
Level 61
Report
you need a great idea, that is how it was solved in chess and go, unless you have that i cant see anyreason why it should do well
OpenAI on the Multi Day Ladder?!: 9/9/2017 04:07:52


knyte 
Level 58
Report
MD ladder is probably significantly tougher to solve with a single agent than the official ladder (with just one template, it's a lot easier to encode things- otherwise, I think you'll need your input representation to encode the map/settings as well or alternatively maybe try some crazy image recognition thing where you hope it's able to figure out the map by itself).

I think a competitive recurrent network-based agent is probably viable with current technology, although it probably won't be able to consistently beat ladder #1's. I don't know enough about other approaches (incl. Monte Carlo tree search, which is what AlphaGo used) to comment on their applicability to Warlight.

That said, while state-of-the-art AI is currently able to do some impressive things, it's worth noting that it's still more restricted than you'd expect. If a human figured out how to be really good at Go, for example, you'd expect them to be able to do a few other things that require (for humans) less intelligence- but AI doesn't translate that way. That's because, at the end of the day, all we're really doing is cleverly finding mathematical operations that seem to work well for a given task. AI doesn't think, or even really learn (at least not in a way similar to humans- neural networks are modeled after our brains, ofc, but current learning algorithms aren't very similar to how you and I learn, and some things that are basic to humans are still unsolved problems in the realm of AI research).

Even some networks that seem to work very well- image recognition, for example- can be fragile. You can make visually unnoticeable changes to an image and completely trick a neural network to classifying it as something else. (see: https://blog.keras.io/the-limitations-of-deep-learning.html).

Honestly, it's easier than you might realize to try building a deep network that does a good job on the WL MD ladder (you probably won't be able to do it rigorously, but I bet there's a few architectures out there that would adapt well to WL), especially if you have the time to tweak some behaviors manually. But actual AI dominance of humans is probably very far away; I'm kind of hoping we figure out better fundamental techniques in the meanwhile.

Edited 9/9/2017 04:08:23
OpenAI on the Multi Day Ladder?!: 9/9/2017 04:20:01


Tito 
Level 45
Report
I think that's inevitable but well, seriously when that happens why would people allow AI to play ladders and compete against humans.

And while I can see AI competitions happening, why would people care about these more than competition between humans?

You can't replicate the salt, the thrill and the adrenaline into a code
OpenAI on the Multi Day Ladder?!: 9/9/2017 04:28:23


knyte 
Level 58
Report
Same reason that AlphaGo was so cool. It picked up on things humans completely ignored. Eventually gaming AI tends to supplement humans.
OpenAI on the Multi Day Ladder?!: 9/9/2017 05:08:26

player12345
Level 61
Report
"It picked up on things humans completely ignored." Apparently the same thing happened with the OpenAI DOTA2 1v1 bot. One top player claims all the pros are now copying the bot:)

The defeat of the top players and the speed at which the bot was created and trained (2 months apparently) came as a huge surprise to the pros. However, DOTA2 is usually played in 5v5 mode, and it features additional complexities. For example, before a game begins the two teams select playable characters in a multi-round draft process. The two teams alternate selections until they each have 5 characters. Each character has a unique set of strengths and weaknesses. Subsequent selections are made to counter earlier ones. The composition of characters determines which strategies will be most effective.

With all these complexities it seems unlikely OpenAI could achieve similar success in any short amount of time. Fortunately, we don't have to speculate: the OpenAI team is apparently working on a 5v5 bot.

Edited 9/9/2017 05:08:44
Posts 1 - 8 of 8