New neural networks coming to PA

Just promise us that you'll stand there with a shotgun pointed at your computer the whole time.

Just in case.

You know, aside from a few edge cases, the PA AI is very, very good. It feels so much more like you are playing an actual (though not very creative) person. It raids, expands, scouts, and attacks at different spots. It only consistently derps against turrets behind walls, where it often will lose an entire superior force because it wont commit to an attack and just charge the dang walls.

I recently bought Grey Goo (which I thought was very fun, had a good campaign) and it's AI in skirmishes, even on the hardest difficulty, was pitifully bad. Only with preset scenarios in the campaign was the AI any good.

I also went back and played some of the old strategy games I have.

Games like Age of Empires at first appear to have surprisingly good AIs, except for the fact that they do the exact same thing(s). Its very easy to force their AIs to do things, which I guess is because they have large scripted responses to things.

Starcraft 2 is even worse- they attack at specific times with a specific amount of units EVERY SINGLE GAME. You can even make them 'juggle' their army. Wait for them to send their blob at you, attack their base, then retreat. Their entire army will run back home, find nothing, then march at your base again. Attack them, and this repeats endlessly.

The vanilla Supcom FA AIs are also pretty sad, and very predictable. Supcom 2 actually has a pretty good AI, it still wasn't necessarily super hard to beat, but at the same time it didn't seem to follow a series of scripts like most of the others. Though that may just be because I abuse the Cybran bomb bouncers, holy crap they decimate air.

C&C Red Alert/Generals etc are pretty lackluster as well. The harder AIs in those also cheat pretty bad to pose any challenge.

Not the most effective approach, IMHO.

Choosing a somewhat genetic approach when training should yield much better results. Don't work with one net, but with a pool of nets and have them train against each others (randomly). None of the typical operations (combine, select, mutate) since they are yielding undesired results on neural nets, respectively because mutation and combination already happen during the regular competition / training.

The end result should reduce the risk of falsely adapting to a local minimum. Oh, and also the vulnerability to bordercase scenarios....

The 3 operations you would usually apply to genetic pools. Recombination (take two specimens, splice and merge them), selection (have them compete in solving a specific task, kill the weakest) and mutation (toggle random parameters to add variety to the genetic pool).

However, doesn't apply to you since you don't want to create a homogeneous species of AI (which happens inevitably if you enforce compatible patterns on the recombination phase), but maximum diversity. By having two specimens compete against each other, you are already coupling them together tight enough.

Your training function already includes a random element, so you can save on the mutation step as well.

Early selection isn't necessary either, since the training function can increase the quality of weak specimens even without killing them.

In theory, a pool size of 5-15 specimens should be sufficient under these conditions to trains the AI against most possible border cases.


For the final selection, just perform a regular round robin style tournament. There's no actual need to limit yourself to shipping the best one, btw. Having support for multiple AI flavors was a requested feature either way.

@exterminans, what have you done?!

(does sound pretty interesting to be honest...)

Lol, i had just watched that one conference where you're talking about neural nets. Completely forgot that it was that old.

edit: And it was about SupCom 2. Always the small details...

Now we're getting a dozen Skynets. That's gonna go well.

Nothing to worry about (yet). The net is far to small and has no feedback loops or storage capabilities, so it's impossible to develop any sentience.

Training in pools like that isn't the most effective method either, but full scale genetic evolution of the net would require hidden layers of 3-4x the width they have now as there needs to ocurre a natural alignment of specific functionalities inside the net which leaves enough spare space to allow for new patterns during the mutation / recombination phases without destroying the established core functions. Just have a look at your own DNA, how much unused trash and inactive genes you have floating around in there, and how much more compact the DNA was if the inactive parts were stripped for the "productive rollout", and how these inactive genes are necessary nonetheless as they still tend to get activated on occasion when placed in certain combinations.

Well, but I'm not going to recommend using broader nets, as that also means that the CPU overhead for the execution of the net in production would explode.



Another thought I had yesterday:
How well is the AI trained towards attacks with joint forces? Does the net have inputs which evaluate the possibility of reinforcements from a different layer? Talking about real simple stuff, like the Boom bot -> Bumblebee -> Dox rotation. (Presence of allied bumblebees changes the target priority for boom bots, instead for anti land defenses or factories, they treat ground based AA as priority targets. Bumblebees aim for anti ground towers to ensure survivability for the next (non-air) wave. Dox come in the 3rd wave and wipe mobile units, especially Infernos, which pose the biggest threat towards boom bots.)

Indeed interesting. There are some things I wouldn't agree with, like the definition of a simple algorithm as intelligent. But this is just bias on the definition side. They just want to persuade you that there is a smooth transition between an autonomous system and a intelligent system. (Being involved with autonomous car development, makes this even harder to believe. Don't get me wrong, I do believe AI is possible, but I would guess it is a lot harder than just pile up some computational power and wait a bit.)

Since I'm a part of the anxious avenue (midway in the the second part) and I "like" the story of Turry, the AI that ,in the mission to get a better hand writing, turned our galaxy into a heap of friendly notes. Both of your arguments don't make me feel better.
We know that everything that can be done and that gives someone an advantage over someone else, will be done. History is full of those examples, and most of them aren't high points of humanity. It is not a question if it going to happen, but when it will happen. Maybe some of the AI projects will not succeed the first time, but there are to many advantages that can be gained, to not try it again and again.

/offtopic

But what if that's just what they want you to believe?

Great to have you back @sorian! This is definitely an exciting (and scary) improvement.

This new tutorial seems relevant! From perceptrons to deep learning. Maybe PA could start using deep nets?

http://www.toptal.com/machine-learn...s-to-deep-networks#snag-only-knowing-devs-now

Wasn't that what the blog post was about?

When the AI puts it's DOX in a formation that spells out "Time to die"..

Hahaha I would be so happy. I hope the ai will start sending troll messages right as it nukes you or smashes your rock with another rock. As the annihilazer powers on it sends a smiley face and game over noob in chat.

Wooooo .... I've rarely seen so much stuff explained in such a smart and easy way to understand. Thanks for sharing this global picture which helps a lot before jumping into the big swimming pool

http://aigamedev.com/open/editorial/2014-awards/ what is this websites freaking issue?

they're a website all about game AI and they NEVER awarded sorian? not even PA's current AI??? what ?

Just saying...

remember, community vote, it's not objective at all, just a popularity contest. I'll take the editors pick any day over that.