Reforged AI for computer?

Hello! Do you know if they’re going to rewrite the AI for the computer (just like in Age of Empires HD) or not?

8 Likes

I am guessing there will be some minor AI improvements, especially since the balance patches are changing quite a bit so hard coded build orders or army comps may not longer be valid. However do not expect anything too much better, cheese strategies will probably still stomp them.

1 Like

Would love for them to put in some machine learning that makes the AI self learn (people have done this for Chess, Go, etc.) but I’m guessing that is a step too far. :stuck_out_tongue:

3 Likes

I hope so. The current ai is too basic and predictable, it just builds and trains the same units every time.

There’s a fan made ai (amai) that is more entertaining and difficult to play against, as it builds and masses different units, and it also buys objects and uses tavern heroes, but unfortunately it doesn’t work with the current patch.

1 Like

I think the normal AI is great for beginner’s practice - it tries to expand and trains a fair amount of units.
Tweaking it according to the current changes definitively sounds like a good addition to Reforged.

make insane AI, use warpten + greedisgood ( AI already has 4x income). And there u go non stop 100 food army fights, which are challenging enough to train ur micro and macro.

We are still at least 3 years from this becoming common. Only top of the range consumer GPUs such as NVidia RTX 2000 series and modern Apple products have hardware neural network acceleration. Intel has it planned, but it will probably be another 2-3 years before the chips hit the consumer market in a big way. No word from AMD of such acceleration happening soon.

That is completely different. I’m not talking about AI machine learning on the client. I’m talking about machine learning on the server. Should be the same learning AI for all, not client specific. That is definitely a few years away at a minimum.

2 Likes

i hope they make IA harder to play campain at imposible level

1 Like

There is no server… Everything the AI does is decided by each client in a game. Hence it has to be computed by the client CPUs and hence it is limited to common consumer hardware. Without common consumer AI acceleration it will likely tax CPUs too hard, especially budget ones, that it will seriously impact performance.

Campaign AI operates differently from normal AI. It purposely is stupid as it often starts with a huge advantage. Campaign difficulty usually comes by adjusting how much of an advantage they have or by making the mission requirements more strict.

I don’t think you are understanding what I am saying. The AI is only calculated at Blizzard. Updates to the AI code are then distributed to the clients through updates. It’s really not that complicated.

One problem is training neural networks. Another problem is running an already trained neural network. Especially good ones with a decent scale, as would be needed by RTS AI, are computationally intensive to run in realtime.

Below is a paper that is slightly dated but does cover the problem to some extent.
https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/37631.pdf
It is worth noting that GPU acceleration performance is purposely reduced on most consumer level cards from NVidia as part of their product family approach to force AI researchers to buy high end, expensive, Quadro and other such industry grade cards. Even if the card could efficiently accelerate the neural network it would be eating up the GPU’s time which in turn could impact frame rate. The exception being the brand new, expensive, RTX series of cards which are consumer grade (rich consumer…) cards which have dedicated tensor cores to speed up both training and execution of neural networks with reduced impact on standard GPU work loads. On these RTX cards currently some games like Final Fantasy 15 use a pre-trained neural network for a mode of AI driven anti aliasing.

Modern Apple products also have such dedicated neural network training and execution acceleration hardware. However again this is not really a product you can expect most people to have.

Maybe you were referring to machine learning to write better traditional AI scripts? This would basically be AI writing code, which might be a thing in cutting edge research but is likely beyond a normal game company.

I’m talking about stored neural networks based on server-side computation. This is for a game, not a complex issue, like how to be a human. It really shouldn’t be a problem running them.

If it were so simple then why are no RTS games doing it? Why is it taking millions of dollars to try and make AI to play DotA2 and LoL at high skill play levels? They have 1v1 working but are nowhere close with team play. I suspect you are underestimating how much of a complex issue real time strategy games are.

Due to the complexity of parameters to feed in I would imagine quite a large neural network (in terms of number of elements) is required. Even though running a large static neural network is a lot faster than training one, it is still not computationally trivial for a CPU. Especially considering that up to 23 of them could be running at the same time (one for each player). Without dedicated hardware acceleration I do not see this being viable, especially considering people expect it to run on older or cheaper hardware.

For now I think purely throwing a guy to work on the AI and AI support systems for a few months would be more worth while and yield AI that performs considerably better than current without being that much more computationally intensive. As nice as neural networks are, I do not see them being common place in games for at least a few more years and AI hardware acceleration is common in consumer systems.

We know that the melee AI is going to be the exact same, considering that Warcraft III: Reforged is going to be 100% compatible with the classic version of Warcraft III.

The only tweaking I can see Blizzard doing concerns the AI for the Warcraft III: Reforged campaign maps that get overhauled.

They cant even debug 1.30.2 . There’s no way they can implement your idea on the server side, and i hope that they don’t read your comment.

1 Like

We do not know if AI will be upgraded or at least revised between now and then.