Could we make T4 dwelling and summon only please ?

[sikbok]

Hi everyone,


On a NN based AI. Please keep the discussion civil.
Its a pretty complex topic, so while it boils down to a 'will this work now [true/false]' we need a whole lot of text more to get to the reasons why it will, or won't.

If someone starts a new topic on this I'll chime in there too.
I made some NNs before I got into game development, so I hope I can add to the discussion.


On T4s - and back on topic - I know this was a thing when AoW3 had just launched.
However we did address this in subsequent patches, so I wonder if this is still a thing that happens in the current game.
Has anyone run into this issue recently? If so, can you tell me a bit about the circumstances (number of turns played, number of players left, are those players humans or AI, etc.)?

Then for Planetfall. We changed a couple of things - not unit tiers though - that will have quite an impact on how the game plays.
While not directly addressing T4 amounts, I wonder if this will actually still be an issue in Planetfall.

 

[sikbok]

Analyzing the screenshots we have seen I assumed the last "resource" was a T4 cap "resource". Ref:

Speculating now but I imagine the resource being from left to right: Energy, Ore, Research, Influence, Morale/Happiness, Operation/Command points, Tier 4 cap.

For starters, your speculation is a bit off. However, we do have a 'fancy resource' somewhere there in the top bar that does something close to what we are discussing : D>
Calling the fancy resource 'Tier 4 cap' doesn't really do it justice though. First of all, the resource is used for high level units - not just T4s - in the current setup. In addition, Planetfall has unit mods which can be applied to all sorts of units. That - together with production rollover - will change the power balance between tiers quite a lot.

And to address some of the comments about T4 design. We've learned from AoW3 and took an angle for Planetfall that should resolve some if the issues noted in this discussion.

Players have enough reasons not to spam T4 Mono-stacks but AI King and Emperor does because they basically don't care about upkeep. This still happens, especially with rogues and warlords. Their army is about 50% T4 and they put their T4 in offensive Mono-stacks. Players would only occasionally face stacks of other things. This happened last weekend with 4 player FFA, medium map (2 layers), emperor difficulty, city founding off, city density low, unifier and seals off.

Hmm @sikbok there you have your recent issue... But I must say I haven't had an issue with the AI spamming T4 but then I play multiplayer mostly so I wouldn't know. @Nerdfish do you play the patched up version I thought one of the patches included something like "AI will not build monostacks ever".

So, from this I take the issue is not T4 production but mono-stacking. Given the classes this happens for - Rogue (Shadow Stalker) and Warlord (Manticore) - I'd assume this is an emergent behavior of how the AI plans its missions and moves units around. Given the movement types - both non-walkers - they'll likely get clumped together naturally.

Do wonder nerdfish, how many turns in did this issue occur for you? The 50% T4 sounds on the high side. It might be due to unit retention and T4 being more survivable. It might also be due to a few cities with very high production, which is where the question about turn count comes in.

 

[Chaingun]

As I said before, you'd need more games than have been played of AoW3. Training set required is to large, and the amount of input data is to many.

Image classifiers are trained using standardized databases. The same approaches can work here having the AI trained on replay databases. It does not require millions of matches, because it can treat every turn as an independent board state. If a single game last 50 turns then it takes 20,000 games to reach a million data points. it should be a simple matter to wait for that many replays to be uploaded to the database. "Opt in telemetry" as Brain pointed out, could provide that many games possibly in one month. (100 games from 200 players each).

As for actual coding, I have being writing AI in tensorflow for the last year. The lion's share of work is designing a representation for the game world (input layers) and make sure gradient flow from output to input. It'd take years for one person, or a couple years for 2 person, or one year for a team.

There is nothing fundamentally insurmountable. The real issue is that it's expensive and people are lazy.

There's no economics behind it. Consumers don't have super computers, and development plans don't have empty spaces of time where games remain stable enough to do this kind of research.

For it to take off, there needs to be a middleware offering that does not kill consumer hardware (or runs in cloud, in the cases where that's tolerable) and that developers can be confident can solve their problems with low uncertainty.

You're possibly overestimating supervised learning. You can train a classifier where each classification makes individual sense but where the policy of following each independently classified "best move" in sequence is abysmal because the adversary is actively studying and responding to what you do and will do their best to mess with your fixed policy. You need reinforcement learning for these kinds of games.