He will be speaking about his use of neural nets in supcom 2, when Is the GDC 2012 summit is it gonna be streaming somewhere? I wanna watch it.

GDC 2012 began on Monday and will continue until Friday.

Ah, I see, thanks

He gave his presentation earlier today, if Farcebook informs me correctly.

Hopefully he'll have a video to share, or at the very least a nice new blog to post. *fingers crossed for juicy technical details*

Why wasnt this plugged more?

Also: DO WANT VIDEO

Yes, either a new blog post or a video would be great

I will be posting to my blog as soon as I get a chance. I will also post the pdf of my slide presentation. It just wont be the same as being there.

I don't have a video to post, however, and I can't just copy and post the GDC vault video.

Thats great! a blog post + slide presentation Is all we need

I'm happy with that.

Cheers, Sorian.

This is awesome, I can't wait to read it.

Thanks!

Here is a preview slide. I hope to have the blog post up this weekend, or early next week.

I'm looking forward to the rest

No more than eight words to a bullet point and injected humor.

Awesome.

For Sorian, a neural network is the same as a honey badger.

I hope to have the blog post up tonight.

Here is one more teaser till then:

Blog post is finally up: http://soriandev.blogspot.com/2012/03/gdc-2012.html

Great presentation! I have always been impressed by the performance of the AI. Would it be possible to have an AI adjust to player tactics, or would that require a lot of processing power/disk space? For example, it was easy to trick the AI by using micro, it would simply crash its armies into you thinking it could win. Would it be possible to have an AI adjust to use tactics that appear to be more effective against players, or would the required dataset for such decisions be way too small? I remember hearing stories about you having AIs run battles on max speed 24/7..

If I understand you correctly, it would basically require that the neural networks learn in the field. This is definitely possible and I did bring it up, but that feature was never approved.

I was also going to experiment with more tactics (rather than just who to target) for KnC.

Would it be possible to design a neural network that learns against a specific player, from one game to the next? An AI that that is tied to your profile, and that with each successive game adjusts itself to better counter your specific style of play and tactics?

Possibly. I am pretty sure we could tie it to your Steam ID.

(Haven't had a chance to check the slides yet, but)

One thing I thought about with the AI Overlord is how, for example, it can be a very powerful tool when in a 2v2 against two AI along with an AI friendly. The two enemy AI are basically combining forces to create one larger entity, but the same may not happen on your team unless you coordinate your every move with your AI teammate.

Have you ever thought of creating an Overlord that would work along side a human? For instance, it could be communicative (whether it be auditory and/or visual) as to what their next move(s) or plan(s) will be. So the AI teammate may say "I am pushing forward with my group of tanks to here!" and will cue the human to the position. Or, perhaps take it one step forward, and give the human the ability to call specific attacks, tell the AI to cancel or retreat their plans, or coordinate in some fashion.

Because nothing is more difficult to have your AI teammate send his tank blob into the combined enemy's blob force and take massive losses. Or trying to defend a two-player attack by yourself. The AI Overlord impresses me and ramps up the challenge, and I wonder how much further it could be developed.

Some games such as CoH have an attack/defend/point out "feature" with the AI, but they are so rudimentary and crude that they don't work as well, or at least with unintended consequences.

Yep. this feature was talked about, but never had the time to implement.

Very interesting read!

Sorian, I have a question about how the ai defends against an attack? Does it use a platoon? looking through the files quickly I can see the land, structure and air platoons, wich each dictates a certain number of units to form an attack platoon (In the PlatoonTemplates).

If I amp up the number of units to a platoon, so it first attacks with a LandAttackTiny of 50 units instead of 3 units, and send a squadron of good ol' loyalists behind the enemy base, the ai doesent use its units to defend: instead builds up its 50 units then moves out to attack while I destroy the base.

How does it take on defending, can you give details on it?

Well, sort of yeah.

My idea is, when the AI makes a decision to attack, it records the situation and the result, and collects a dataset over time. After the dataset has reached a certain size, you can then see what unit mix performed best against X, or when the AI's decision reliably failed, and adjust these strategies in the future. This would basically require gathering it from a large set of players, because otherwise the results could be off due to a small dataset.

To take it a step further: I dont know if its possible to make an AI distinguish between different levels of player skill, because of course results vary depending on player micro, and you could then further adjust the AI to either the opponent, or make the hard AI make different considerations or something..

As for using different tactics: What the AI does now is try to find a weak point and attack, correct? What sort of tactics would you like to add? Deliberately make platoons for the sole purpose of harassment, killing buildings, etc?

Do we really need more bots giving harassment out i though mav was the only 1 qualified for it

sorian nice any change of my mid going boom now after reading it all lol
love this part