Training the AI to build better ships

[ZorinW]

I just had an idea for how to gradually improve the AI's ability to build ships. The devs should implement a feedback form in the post battle screen. So you can select each enemy design and rate it from -3 to +3 on a set number of parameters. Say number of guns (mains/secondary), armor, offsets, speed, smoke (number of funnels), number of torpedos, etc. 

Over time the AI will then "know" what makes a good design.

What do you guys/gals think?

[PhoenixLP44]

Interesting Idea but implementation is a whole new level and would take way too much valuable development time.

[ZorinW]

24 minutes ago, PhoenixLP44 said:

Interesting Idea but implementation is a whole new level and would take way too much valuable development time.

But in the end the enemey ships we are fighting are the single most important part of this game. As it stands right now, there is absolutely no challenge in playing the campaign. I never lost one and you are even going to win when effectively binding your hands behind your back because the AI is stil so bad at coming up with ships that would pose an actual threat to players.

[kjg000]

Excellent idea but I suspect it would be impractical. It would need some expertise in deep learning and a lot of coding I think. 

However it does suggest an interesting experiment, have a playoff against, say, 10000 ai designs were only the winners progress to the next level and see what the 10 best designs are. Not likely to happen but would be interesting. 

Hmm, I wonder if the Devs would consider (eventually) setting up a comp where user designs are allowed to compete against each other under ai control? Could be worth a YouTube channel. 

[HistoricalAccuracyMan]

1 hour ago, ZorinW said:

Over time the AI will then "know" what makes a good design

The problem with that statement is that it is highly subjective, and very dependent on the ship in question and the consequences surrounding it.

You can design a ship for a specific role or purpose, but at the end of the day...what makes that ship a "good" ship is whether or not it can fulfil it's role in a satisfactory manner. I might say that a "good" BC should be able to outgun what it can't outrun and outrun what it can't outgun. Somebody else might say that a "good" BC should be a cruiser's worst nightmare and give BBs a run for their money in a stand-up fight. Who is right? One of us? Both of us? Neither of us?

The problems I see with a "rating" system is that it has the potential to be "gamed" in a sense, or at least, not used in the way you probably envision it. Let's just say that you rate a certain AI design as -3 on "number of guns" because you think it has too many. If you just keep doing that and all the AI "knows" is "oh...looks like I'm still using too many guns. Looks like I better reduce them again," I fear that you have the potential to encounter an enemy ship that is the equivalent of a 1890s battleship armament (two main gun turrets, handful of secondary guns or lots of small guns) while you have a 1930s super ship with 18" high velocity long barrels and 8" secondary guns. Or even worse...the AI "learns" that "Guess I'm using too many guns....let's make a 15,000 ton CA a high speed torpedo sled!" and now, you are doing nothing but dodging walls of 40 torpedoes at a time and you can't chase it down because it's doing 36 knots or more. Yes, they could program in some checks/balances to make sure that players can't just game the system to "force" the AI to design subpar ships, but that would take away a lot of valuable time that they could use to work on/improve other things.

I personally think that if the AI were to "loosely mimic" our designs...that might do some good to at least "improve" the challenge the AI presents (when it comes to ship designs and not tactics/maneuvering anyways). How it would work is basically like this:
The AI notices that you currently have or are building BBs with 14" guns, Krupp II armor, double geared turbines that travel at 24 knots. The AI then looks at it's own fleet and says, "Yikes...I've still got BBs running 10" guns, Harvey Armor, and barely making 19 knots. But I've also got some BBs that are sporting 12" guns, Krupp I armor, and can make 22 knots...let's change that." Then, the AI scraps the "obsolete" designs and either decides to retrofit their "better" designs or build new ones that are more or less on par with yours (as technology levels will allow). I'm not saying they'll make carbon copies of your ship, but the AI might turn out a new BB that has 15" guns, Krupp I armor, turbo-electric drives, and 23 knots of speed (i.e. something that is decently/competitively close to yours...but not overpowered or severely outclassed). Of course, this would spread to all nations, with some variety mixed in so maybe you do get the wacky design or two since it is the AI after all.

I'll also re-iterate that I'm fairly certain that the devs have mentioned something like a "ship design library" that the AI would pull some designs from, where either pre-defined or user-submitted designs would be entered (however, I think they only really said that would apply to the "historical" AI behavior so you could actually fight against "historical" ships). Whether or not that will happen, I honestly don't know. I really hope it does and I hope it would extend to both Random and Historical AI behavior because if you think about it...how many different designs have probably been thought up by the community as a whole--both "historical" and our own original creations--that are either solid ship designs or would be fun to fight against?

[neph]

You could set up an adversarial neural network, let the AI play custom battles against itself for a couple months, and then use those results.

[PalaiologosTheGreat]

You could of course just make the AI just use historical designs/upgun them. That would be pretty nice. 

[TiagoStein]

On 9/22/2022 at 1:51 PM, kjg000 said:

Excellent idea but I suspect it would be impractical. It would need some expertise in deep learning and a lot of coding I think. 

However it does suggest an interesting experiment, have a playoff against, say, 10000 ai designs were only the winners progress to the next level and see what the 10 best designs are. Not likely to happen but would be interesting. 

Hmm, I wonder if the Devs would consider (eventually) setting up a comp where user designs are allowed to compete against each other under ai control? Could be worth a YouTube channel. 

Not so much ( I work with  deep learning) and  this would be a rather simple model. BUT they NEED to finish ALL the  shipbuilding of the game before they do that otherwise any effort if wasted. You need  the exact mechanics and values set so one can define the input and output forms of the model, then you  can even make a GAN where 2 neural models  dispute with each other and learn themselves to build  good ships.. But a warning...   they will bcecoem so much better than players that we will not even bother to builg our own ships.

 

[Wowzery]

On 9/22/2022 at 11:43 AM, ZorinW said:

But in the end the enemey ships we are fighting are the single most important part of this game. As it stands right now, there is absolutely no challenge in playing the campaign. I never lost one and you are even going to win when effectively binding your hands behind your back because the AI is stil so bad at coming up with ships that would pose an actual threat to players.

Some people the devs will never please.

And this is why I auto-resolve all battles in the campaign, so I don't feel like a cheater.

[kjg000]

On 10/12/2022 at 10:32 PM, TiagoStein said:

Not so much ( I work with  deep learning) and  this would be a rather simple model. BUT they NEED to finish ALL the  shipbuilding of the game before they do that otherwise any effort if wasted. You need  the exact mechanics and values set so one can define the input and output forms of the model, then you  can even make a GAN where 2 neural models  dispute with each other and learn themselves to build  good ships.. But a warning...   they will bcecoem so much better than players that we will not even bother to builg our own ships.

 

It depends on what you want to achieve, I think watching what the AI built would be helpful in refining the shipyard and deriving better and more realistic design limits without resorting to a lot of arbitrary rules and restrictions. It would also highlight defects in the shipyard such as floating secondaries etcetera.