Improved Pro AI, please give it a try and post feedback

[LaFayette]

As of today, a number of fixes landed for the Pro AI and there is a new Hard (Beta) AI that is even smarter. The new Beta AI tends to beat the Pro AI handily as axis vs allies, and puts up a better fight when playing as allies. Check it out, download the latest TripleA and give feedback here!

I found the new Beta AI is still not great, but puts up a much more interesting fight!

Ask:

• download the latest TripleA
• in engine settings > testing > turn "beta features" to on
• play a single player game against the Hard (Beta) AI
• post your savegames here with feedback, mention too which version you ran. For feedback, looking for general commentary, play-by-play, blunders, things that were good, bad, opportunities missed. This feedback will go directly to improving the AI play.

Fixes to Pro(AI):

• 3:2 Places in sea zones opened by placementAnySeaZone (#15148)
• 3:3 Lets placementCapturedTerritory players use fresh conquests (#15144)
• 3:4 Purchase sim keeps the attackers of undefended battles (#15141)
• 3:5 Counts capital income only when the capture stops it (#15142)
• 3:6 Counts a captured capital's income once (#15143)
• 3:7 Keeps searching for raids past bombers already attacking (#15131)
• 3:8 Restores the non-combat sea zone values (#15127)
• 3:9 Counts free carrier deck space when buying for a sea zone (#15126)
• 3:10 Skips production in territories the purchase sim conquers (#15123)
• 3:11 Keeps a land unit alive when the capture is worth it (#15128)
• 3:12 Retries amphibious targets that were removed for an attack later dropped (#15125)
• 3:13 Purchase sim no longer captures raided or air-only targets (#15122)
• 3:14 Capital defensive stance uses the copy's move map (#15124)

New Capabilities of the Hard(Beta) AI:

• 2:2 Adds the Hard Beta AI behind the beta features setting (#15129)
• 2:3 Defends a capital that is about to fall (#15149)
• 2:4 Uses bombers an attack doesn't need for raids (#15132)
• 2:5 Lands units behind a won sea battle, Axis only (#15134)
• 2:6 Stops buying warships that sit idle (#15139)
• 2:7 Places bid units for round-1 battles (#15136)
• 2:8 Values territories by national objectives (#15137)
• 2:9 Scores land purchases as defenders (#15138)
• 2:10 Buys transports only when cargo is waiting (#15140)
• 2:11 Strafes land attacks that can't be held (#15135)
• 2:12 Blocks enemy amphibious lanes with a warship (#15133)
• 2:13 Leaves production slots free for held units (#15130)

New updates in: 2026-November-01 - 2026-11.16.15866

For Hard(Beta) AI:

• subs will now reinforce navies if they can submerge before battle and if they can't be sunk due to missing destroyers. This is useful as the sub becomes grouped with an existing navy, avoids leaving stray subs, and the extra sub forces an attacker to send a larger strike group.
• "extra" planes will now be sent to the highest value battles that are not yet guaranteed rather than being first sent to the lowest odds battle first.

[LaFayette]

Some fun notes, I was able to download everything ever posted in the forums and scraped out a list of every complaint ever of the AI! I used that and the PBF games to help resolve weaknesses. I set up little scenarios and called them kata's that the AI would train on. Then had game after game run to evaluate improvements (32 cores at 5GHZ running TripleA games non stop for almost 4 days).

Really interested to hear how this updated AI does, and any detailed "vs AI" game write ups would go directly into more improvements!

[ubernaut]

@LaFayette nice will check it out

[RogerCooper]

@ubernaut When I download I see "Hard Beta AI" and "Flow Field AI". What is "Flow Field AI" and where the "Pro AI"?

[ubernaut]

@RogerCooper i think this was for @lafayette

[RogerCooper]

The game does not load if "Flow Field AI" is used?

[LaFayette]

Hopefully this clarifies which is what:

Hard(AI) = Pro AI

• now includes several important bug fixes

Hard Beta (AI) = Pro AI with beta features (this is brand new)

• has all the same bug fixes
• has additional features, for example, aware of national objectives

Flow Field (AI) = A totally different beta AI

• dates from about 8 years ago, a promising new AI that is at early stages. It not being functional is known.

[Cernel]

I just launched "World War II v5 SE TR" and saw that, on the first Russians turn, the Russians submarine is not moved to 7 Sea Zone. The submarine should go there to reinforce the British fleet even more so on the account that Axis do not have any destroyers anywhere nearby. This way, the Russians player is free either to use the submarine to defend or to submerge it if the Germans attack with massive forces (meaning that the AI should not submerge it if the Germans attack with forces that are not too powerful so the submarine may significantly increase the Germans losses and submerge it if the Germans attack with so many forces that the submarine would otherwise do little more than being destroyed). The next round, the submarine would also have more options, like attacking a surviving Germans transport in 5 Sea Zone (saving the British from having to dedicate a fighter to it).

On round 1 in that game, the Russians submarine should go to 7 Sea Zone every time. When the Germans attack 7 Sea Zone on round 1 (if they do), the submarine should submerge or not submerge depending on the entity of the Germans force attacking.

The matter of the submarine going there would be debatable if the Germans had a destroyer at reach (which is not the case), but I would say that the submarine should go there anyway as there is not much else that it can do if left in 4 Sea Zone.

[LaFayette]

@Cernel would you mind attaching a save game?

Otherwise that write up is perfect! I can capture these commentaries plus save games and create "skill" scenarios to tune the AI player. I mean, go to town with your write-up (just so long as the save game is included)!

I wonder if there should be a dedicated thread for these kinda of posts?

All in all, human vs AI are the most valuable feedbacks. The latest AI is a result of having played about 25,000 games against itself. The key issue is that the AI doesn't recognize its own mistakes and doesn't punish that game play like it should. Humans do though.

[Cernel]

@LaFayette said:

@Cernel would you mind attaching a save game?

Are you sure you want a game-save? It happens right at the start of the game: select "World War II v5 1942 SE TR", assign only "Russians" to the AI, and start the game. You will then be able to take the game-save after the first Russians turn has ended (in a few seconds).

[LaFayette]

@Cernel Yes, very sure. I want posts that have save-game & write up.

Why?

• I'll have the exact scenario/game and won't be guessing
• I can pull both save-game and forum commentary in an automated script, the feedback is analyzed automatically and then I'll iterate to distill the feedback down into a scenario that can be used for benchmarking and/or generic bug fixes.

FWIW, I was able to pull every forum post via scripting and in theory this latest AI tried to patch every single complaint ever posted about AIs. The set of complaints, the feedbacks, human vs AI games - those are all a lot more valuable than having the AI play against itself to improve.

[Cernel]

@LaFayette This is the game-save:
2026-11-1-World-War-II-v5-1942-SE-TR.tsvg

I'll also point out that, for the remaining part, the game-play is not very good: the AI is taking too much risk and not focusing enough. It should have sent everything (EVERYTHING) it can send to Ukraine SSR and everything else or almost everything else which can reach West Russia to West Russia.

I guess (once the max is sent to Ukraine SSR) sending 2 infantries and 1 artillery to Baltic States and all the rest to West Russia would also be acceptable, but I'd say better only attacking Ukraine with everything possible and West Russia with next everything possible.

In any case, for sure Ukraine SSR must be attacked with max possible. That is out of doubt.

[LaFayette]

I'd love to see extended write ups even. A save game with every blunder pointed out, every flawed strategy, and any praise for what the AI did would be perfect. The analysis step of the write ups is almost entirely automated.

Those write ups would lead to quick and substantial improvements to AI play.

The AI play against itself leads to very small improvements, usually bug fixes and not deep strategy improvements. These human vs AI games and detailed write up is (at the moment) by far the best way to further improve the AI. Since the analysis step is largely automated, if 50 writes up landed today, the AI available tomorrow would likely be several echelon's stronger.

[Cernel]

@LaFayette About the attack in Ukraine SSR, the only actual difference is that both fighters would go there (as part of sending there everything which can reach). The AI actually sent everything at reach except 1 fighter (which was instead sent to Baltic States).

[LaFayette]

Perfect, thank you @cernel ; keep the feedback coming!

PS: favor 'edit' to add additional notes on the same save game, so that one post and one save game match together. It keeps the analysis a little bit easier.

[LaFayette]

@cernel a patch is being built - check the latest in about 10 minutes from now.

edit: patch is built and available (Hard Beta AI)

https://github.com/triplea-game/triplea/pull/15159

The algorithm change identified:

• the Pro AI was sending spare fighters first to the lowest odds battle
• the update, Pro AI now sends excess fighters to the highest value battles first that are not yet certain wins

So, Pro AI will now ensure the high value battles first. This update changes the behavior to generally send figs as you described 90% of the time.

[LaFayette]

@cernel - Sub behavior now updated from: 2026-November-01 - 2026-11.16.15866

The sub should be moved 100% of the time now.
The previous release had the fighter fix movement fix.

[Cernel]

@LaFayette I have not tested the new one yet.

I don't know if this may be of help, but I wanted to make clear why attacking only Ukraine S.S.R. and West Russia the way I described is better than what the AI did.

In the latest game-save which I've posted, the AI makes 3 attacks.

territory: forces sent: average TUV swing + %Win * territory production value = average TUV swing + average production increase = total bonus
Ukraine S.S.R.: 3 infantry, 1 artillery, 3 armour, 1 fighter: 12 + 71% * 2 = 12 + 1 = 13
West Russia: 5 infantry, 2 artillery, 1 armour: 8 + 94% * 2 = 8 + 2 = 10
Baltic States: 3 infantry, 1 fighter: 3 + 89% * 2 = 3 + 2 = 5
TOTAL: : : 23 + 5 = 28

Now, the first most obvious mistake we can see is that 1 infantry is left in Karelia S.S.R. during Combat Move.
That infantry could have gone to Baltic States or to West Russia to attack.
Baltic States is questionable because it is highly subjected to counter-attack.
West Russia is not highly subjected to counterattack regarding the average result with the infantry added and the territory conquered (which is 6 units left if the attacker won, specifically 3 infantry, 2 artillery and 1 armour). Assuming Ukraine S.S.R. taken, the Germans on the next turn can send to West Russia up to 3 infantry, 1 armour, 5 fighters. The battle between these forces and 3 infantry, 2 artillery and 1 armour is 8 TUV swing, which is much less than what those forces can do if used elsewhere, which means the relative risk is virtually inexistent.

These are the statistics same as above but sending 1 more infantry (from Karelia S.S.R.) to West Russia

territory: forces sent: average TUV swing + %Win * territory production value = average TUV swing + average production increase = total bonus
Ukraine S.S.R.: 3 infantry, 1 artillery, 3 armour, 1 fighter: 12 + 71% * 2 = 12 + 1 = 13
West Russia: 6 infantry, 2 artillery, 1 armour: 9 + 98% * 2 = 9 + 2 = 10
Baltic States: 3 infantry, 1 fighter: 3 + 89% * 2 = 3 + 2 = 5
TOTAL: : : 24 + 5 = 29

Instead, sending everything which can reach to Ukraine S.S.R. and the everything remaining which can reach to West Russia is the following.

territory: forces sent: average TUV swing + %Win * territory production value = average TUV swing + average production increase = total bonus
Ukraine S.S.R.: 3 infantry, 1 artillery, 3 armour, 2 fighter: 18 + 89% * 2 = 18 + 2 = 20
West Russia: 9 infantry, 2 artillery, 1 armour: 10 + 100% * 2 = 10 + 2 = 12
TOTAL: : : 28 + 4 = 32

This is a total of 5 average TUV swing more and 1 average production increase less for a grand total of 4 more than what at the game-save.

This is per se only a small difference, but it is something.

The other main element is that sending to Baltic States is just not worth it because there are many Germans units which can attack there and nowhere else, so that is a very easy counter-attack for Germans with almost certainly no counter-attack for Russians thereafter. That may be worth it if the Russians would have had spare air units, but this is certainly not the case as they are all required in Ukraine S.S.R..

[LaFayette]

@Cernel Useful feedback, I'm toying around with the idea of "don't create targets for your opponent". A first metric I'm thinking of is "how much firepower could the opponent bring to bear". That's a good single sum. It's kinda complex though how to weight that signal. I'm instrumenting the AI (for benchmarking purposes) at least to capture some data around that metric and when it is clearly useful to consider.