Investigation of AI reassessment upon reload

[therother]

Over the past month or so, a new game mechanics issue has arisen: does the AI reassess its actions after a reload of a saved game?

This thread is for investigating that issue, with the eventual aim of understanding the observations already made by many of the community. To be most effective, especially reading some posts from CA staff members, the thread must be as concise, relevant and conclusive as possible. Please try to be as complete and accurate as you can be before posting.

Also, take some time to familiarise yourself with the posting guidelines for this forum. Any post, or part of a post, which is not related to research of this issue will be removed from this thread.

Edit: In answer to Pode's first post, the sort of things we are looking for are: detailed observations, repeatable experiments, community attempts to gather large datasets, analysis of that data, and hopefully possible workarounds.

Which is a very long-winded way of saying, yes, that's fine!

[therother]

added by A.Saturnus

It seems the discussion around this issue has come to a halt now. Therefore I post the conclusion I have drawn later in the thread here. The participants of the discussion have more or less agreed in general to this conclusion. Only the wording has drawn objections. But for simplicity I will leave it as it is. The objections can be found later in the thread.

• It is certain that after loading a savegame, the AI will most of the time break up ongoing sieges in cases where it doesn't do that without save/load.
• It seems that the objective to take the previously sieged city is not lost entirely, as the AI will usually try to reengage the siege the turn after the load.
• As a consequence, saving and reloading will affect the development of AI factions, at least in the beginning of a campagne.
• It is not yet clear to what extent the course of a campagne is influenced in an undesirable way by this. AI factions do conquer new territories at a slow pace, but it is not clear whether this is due to the save/load issue or general weakness of the AI. There are three possibilities:
  
  • Above the weakness of the strategic AI, this issue is irrelevant as its effect is only noticeable under extreme conditions.
  • Saving and reloading often does noticable affect gameplay, but doesn't make the game unplayable.
  • The effect is so pronounced that it can be called a "game-breaker".
• A clearly noticable aspect is that, when saving and loading often, AI factions will not take the opportunity to capture rebel territory to the same extent as it does in continued playing.

It should be clear that point four is the most important aspect here, as it concerns the impact the issue has on gameplay. From reading the evidence in this topic everyone may come to his own conlusion on this.

[Pode]

Not sure if this is the type of post you're after here, therother, but this was what convinced me.

RTW 1.2, mods to build file only, M/M campaign as Britons, toggle_fow off. Zero player activity after first turn, simply hit end turn until an AI laid seige somewhere. When that happened, I hit ctrl-s, then ctrl-l, then end turn. Test ran to winter 265. 13 of 13 sieges were broken off by AI, zero provinces changed hands.

Have not tested the protectorate acceptance implications yet.

[hrvojej]

I'll quote my post from the thread in colosseum where this was discussed (minus the snide remark at the end):

A suggestion for a test:

Start a number of campaigns always with a same nation (one that is regularly unlockable) in vanilla unmodded RTW 1.2. Play a number of turns without reloading and without doing anything. Toggle FoW off, and count the number of provinces that have changed hands. Now do this with saving and reloading every time, and after the same number of turns count provinces that have changed hands. If a few of us do the same thing, take the same nation, play the same number of turns, use the same counting method etc. everything should be controlled for, and we would get a bigger sample size in a less time-consuming fashion. So, we would have to agree on the faction, No. of turns, No. of tests each of us would do, etc., but we would have respectable dataset in a short time, and without any of us spending several nights just clicking end of turn button.

After we gather all the data I volunteer to do the statistical anaysis on it and then we'll have cold hard evidence that something is going on. And no more "probability theory" arguments either - as we would have solid basis to counter them (another reason why we need a large sample size).

If people are interested, I'd be happy to elaborate further. Personally, I'm not sure it will change anything before XP though (if at all), but I'm willing to lend my time to this investigation.

[hrvojej]

Reposting from the Colosseum thread in case not everybody who is interested is reading both threads;

__________________________________________________________________

Ok, so, for those who are interested, here is the setup I would recommend:

- use umodded RTW 1.2; this means no custom mods either; we want everything to be the same and pristine in every experimenter's case so as to make the data cross-comparable
- take the Britons (thanks rcp ), medium/medium imperial campaign
- for the control, play 15 turns without doing anything; this literally means don't touch a thing other than the end of turn button; 15 turns is enough to bring two consecutive sieges that ended in starving the garrison out, I believe, hence I think this is a good number - not too big, not too small
- once 15 turns have passed (you hit end of turn for 15 times, to be exact), toggle FoW off and count the number of provinces that are not owned by the original owners; this means rebel provinces as well; this requires you to know beyond any doubt who was the original owner of the province - write it down beforehand if you think it's going to be necessary, or something to that effect
- for the "treatment" group, do everything like you did it in the control group - Britons M/M, hit end of turn (EoT) 15 times, don't touch anything - with the exception that now each turn you hit quicksave(QS) and then quickload (QL); hence, start - QS -QL - EoT - QS - QL -EoT -... etc. until you have hit EoT 15 times; at that point, turn FoW off and again count number of provinces not owned by the original owners
- if at any point in either your control or treatment campaigns you get attacked, abort that campaign, and start anew; we don't want to interfere at all, even if it's only autoresolve, and hence discard that campaign and instead do a new one fromt he begginging until you have the 5 controls (without reloads) and 5 treatments (with reloads)
- make 5 separate control campaigns (each running from fresh start to 15th EoT), and 5 separate treatment camapigns (again each running from fresh start to 15th EoT); save the campaigns at that point to a regular save, zip all those saves (you should have 10 of them, and this shuldn't take up much space) and keep them on your HD for the time being inc ase we need to send these to someone or verify something
- PM me the results, but keep in mind that I can only have 5 PMs at a time, so if I don't write back saying I got it in a day, PM me again; I'll do the analysis of the results, and post them here; I will not soup up the results, or do anything else to them other than analyze them - I do this kind of stuff for a living, and I know how to keep my personal feelings out of it, believe me


It is paramount that we all do it the same way, so please, if you think that you won't be able to do it in the way we all agree to do it, do not send me your data. I really do not want to deal with falsifications, souped up data for whatever reason, results that were collected through deviation of the basic rules, and similar things, as this would invalidate everybody else's effort as well. Let's do it right, or not do it at all.

This should tell us whether there is something fishy going on or not with saves.If we can't get 10 people (well, 9 other than myself), maybe each of us could do more runs. Say, 8 or 10 per control/treatment. Added bonus to this is that it cannot be said that it is due to a single PC configuration, or having FoW off all the time (which is a cheat and hence not an out-of-the-box game functionality). Please feel free to comment on the guidelines I posted above. If we agree, and there are no further comments until about 6pm GMT Friday, we can get to work, and by the end of the weekend, we should already be able to discuss the results.

Cheers,

[mishad]

Another avenue for testing: can we predict when the "periodic re-assessment by the AI" will happen?

If we could, then we could remember to save only at those points (say, every 5 turns, or whatever). This would minimise the effect of the re-assessment on re-load (since the assessment would have happenned even if we didn't reload).

Problem 1: how do we tell whether there has been a re-assessment? Using broken sieges as our best indicator for re-evaluation, how about: turn FoW off and watch the map like a hawk for broken sieges on every end-turn, log such events, including the AI involved, over a large number of reasonably long (50 turns?) campaigns, and look at statistical patterns (it will have to be statistical, because sometimes there won't be a siege in progress when the re-eval happens, and sometimes the siege won't be broken).

Problem 2: it is quite possible be that (absent re-loads) AI re-assessment is staggered through the turns (AI#1 assesses on turns 1, 11, 21, etc; AI#2 on 2, 12, 22; ... AI#10 on 10, 20, 30 etc.) so there is never a "good" time to save. The test above should show this as separate repeating patterns for the various AIs.

Thoughts, anyone?

Also, we should standardise unit size as well: Britons/M/M/Normal?

[Epistolary Richard]

Another avenue for testing: can we predict when the "periodic re-assessment by the AI" will happen?

I'm not convinced the "re-assessment" only happens once in a while. Surely the AI would perform exactly the same processes between each turn. One of the most obvious examples of this issue is that AI leaves sieges and then goes immediately back to them. That kind of behaviour is not really indicative of the AI reassessing its strategy on reload (per the Shogun post) and then again at the beginning of its second turn. It's more indicative that for the first turn the AI doesn't know what it's doing and then it does a proper reassessment for the second turn.

If this is the case and the AI reassesses its position each turn, it would be useful to know at what point this happened. Of the viable times (beginning of its turn, end of its turn, end/beginning of year) it seems most likely that it occurs at the end/beginning of the year - any other time would not fit in with the available evidence.

A useful exercise would be to use Myrddraal's hot seat mod to halt the turn in the middle of the slave faction, save the game there and reload. The game should then reload from the slave faction's go under AI control, go over the year end, perform the proper re-assessment (hopefully!) and then go to the player's turn. It should be clear enough whether the AI has maintained a siege that on a normal save/load it would have abandoned.

[sik1977]

Reposting from the original Loadgame bug thread at the .com, now in the graveyard. I posted this on 27/02/05 on page 5 of that thread

Na, I am sure the problem wasn't as severe in v1.1. I continued an old campaign from 1.1 after installing 1.2. The AI was doing just fine, specially the Romans. They had made substantial gains, specially Brutii and Julii (I being Parhia never met them in combat till v1.2). After installing 1.2 I couldn't understand why Brutii which was so strong couldn't even take the two rebel settlements it would lay siege to every other turn and then break away. It went on for 50 or so turns. I play very slowly and micro-manage everything, hence quick saving/loading is a norm for me. This was true for 1.1 and never hurted the AI like this.

I did extensive testing after reading this thread. I never jump on the band wagon without first testing something myself. I have too much faith in CA to flame them without good reason (well I still can't flame them, they hold a special place in my heart... hehe). My tests confirmed that AI armies not only lift sieges, it always tends to completely walk away from the said settlement as if it has completely forgotten what it had set out to accomplish. Only to repeat the process ad-infinitum.

I thus played yesterday for 8 hour at a stretch without loading once and because i am quite further down my existing Parthian campaign, I only got some 5/6 turns done. And guess what, the Brutii finally took those two rebel settlements that they couldn't for the last 50 turns. They also laid siege to two other towns. Same with Scipii and Julii who each took one settlement and laid siege and maintained it, to atleast three more. However, after 8 hours, i had to save and go to sleep, and when i reloaded later, the AI had forgiven/forgotten its enemies/goals and walked away from all its previous sieges. When I say walked away, they had moved their sieging armies a whole turn length away from the previously sieged settlements.

Hope this helps.

here is the link

http://p223.ezboard.com/fshoguntotal...rt=81&stop=100

EDIT: Reposting another test i did and posted on 3-01-05 on page 7 of the above mentioned thread, this was while trying if there was any difference with autosaves/quicksaves/regular saves and difficulty settings.

Sorry for all that fuss guys. I just play tested autosaves on VH, and the bad news is that the bug is very much present in autosaves just like quicksaves etc.

I played my existing Parhia campaign for atleast 6 turns without loading and tested on a turn which had atleast 5 settlements undersiege, I checked them all repeatedly (by re-using the same autosave; manually copy pasting it back to check), and every single time the AI lifted all the 5 sieges as i pressed 'end turn', without any nearby threat.

So thats it for VH. Lets see if you guys have any luck on any different difficulty levels.

link is
http://p223.ezboard.com/fshoguntotal...t=121&stop=140

[Pritzl]

hrvojej, your control/test concept is perfect imo. With enough numbers, we can do a simple statistical analysis (t-test?) to confirm that the AI's reassessment routinely results in much fewer provinces changing hands.

Might I suggest an improvement to the test though? Instead of manually noting the province allegiances, maybe a screenshot of the campaign-map mini-window on the bottom left at the beginning of the test and end of the 15 turns would improve accuracy. Furthermore, perhaps proper random selection of who tests with saves and who doesn't is worthwhile to remove any doubts of bias?

Finally, is there thread for participants to sign-up and post their results somewhere on the forum?

[edit]I should read better. 5 and 5 per tester and PMs it is.[/edit]

[A.Saturnus]

- PM me the results, but keep in mind that I can only have 5 PMs at a time, so if I don't write back saying I got it in a day, PM me again; I'll do the analysis of the results, and post them here; I will not soup up the results, or do anything else to them other than analyze them - I do this kind of stuff for a living, and I know how to keep my personal feelings out of it, believe me

You can PM your results to me. My mailbox is bigger.

It would be ideal if FoW is toggled off all the time (including the controls), so every single reassessment can be spotted. However, please add to your results a description whether or not your trials have been exactly as hrvojej has layed out.

[therother]

I don't think we need a control for player passivity, as the player will be inactive in both tests, i.e. it's not a variable. It may well be interesting to try and devise a test for AI response to player actions, but I think it is unnecessary at this stage.

[roguebolo]

Well, then, that entirely discards the relevance of a real game, in which the player is actually making moves, but also loading and saving each turn, which is the way most people probably play. To simply refuse to move at ALL for 20 turns is to simply allow it to examine the starting position over and over again and decide amongst the many possible opening moves, but not to allow it to follow up on those openings because you are giving it nothing to "react to".

Most AI tends to be reactive to the human player -- sometimes called "triggers" in the gaming world. If you don't give it anything to react to, any kind of predictor on how you intend to play, it could very well just bide its time until you do. If a "load game" behavior triggers that particular AI behavior, it doesn't really mean much to me when I'm playing a real game and making moves. Because each turn I give it new information to react to when I make my moves.

[Pode]

For the benefit of newcomers to this issue, let me try to distill the past history of testing that lead us to this point. Not all of these points have been established to the same level of proof.
-Standard test methodology has been as described below. Run for a bit just ending turn and note the number of provinces taken or sieges prosecuted if you had FOW off. Run the same campaign saving and loading every turn and watch things fail to be conquered and the AIs move defensively or aimlessly.
-Quicksaves, autosaves, and named saves have all been tested and found to display the same behavior.
-Loading from the manual interface and loading using continue campaign have been found to produce the same behavior.
-Loading every other turn has been found to catch the AIs in a loop of alternately laying siege on the turn not preceeded by a load and lifting siege on the turn preceeded by a load, indicating that the decisions produced after a load are materially different that normal even in very similar situations.
-Timing of the load within the player's turn if the player is active has been shown not to have an effect.
-Player inactivity with loading produces AI inactivity. Player inactivity without loading produces AI activity.
-Two loads per turn without player activity does not appear to affect the AI's pacifism after loading in my VERY limited test last night. The fact that Roguebolo loads multiple times during a turn and has yet to see a single AI siege lifted in his campaigns indicates that this one needs testing badly.
-AFAIK, RB is the only person who has experienced the bug in testing but not in campaign.

That is the body of fact as I see it. My understanding of the purpose of this thread is to gather statistical evidence of these facts instead of the semi-anecdotal evidence currently available. Hypotheses that can explain these observed facts are probably welcome as well (especially testable ones), although that is the moderator's call. Personally, I'd be particularly interested in anyone that can explain or duplicate RB's results. What makes him special that his AI's work?

[therother]

You may well have a point, but IMO it does not undermine the fundamental reason for this test. We are attempting to test if the AI acts differently in the two cases, where the player is triggering AI in exactly the same way, with the only variable being the saving/loading of the game in one case.

[Kraxis]

rouge the AI factions already react to each other... At least that is obvious from the non-save/load games. That we remove a single faction is of little consequence as the game is supposed to actually feature non-reactionary regions (rebels) and dead factions. One faction out of 21 is not much, and in general only means another 2-3 'rebel' provinces.

From all my games it doesn't seem that the AI factions react any differently to us than to all the other AI factions. We are 'just' another faction. And the AI factions have plenty to react to, each other.

[roguebolo]

Unfortunately, I think you're seriously wrong about that and it was one of the main issues I pondered over when I first played MTW. The faction AIs do indeed seem to act as seperate entities, but there seems to be an overriding AI governing the game in a certain direction for the benefit of the human player. I've always thought this is one of the reasons they've skirted around the prospect of a multi-player version of the campaign game. As long as it is not designed to be multi-player, balancing the relative strength of different factions on the campaign map is unnecessary. It allows the game developers to take liberties to impose situations on the human player. For instance, it would be simple for ANY bordering faction to simply crush the Polish faction in MTW early in the game. But they don't. Otherwise, it wouldn't make much sense to try to play the Polish as a human player, right? So this AI "overseer" makes sure that doesn't happen. I think we get a small glimpse of him through the Pope and the Senate, wherein he tries to direct your actions as well.

After you reach a certain skill level in the campaign game, the AI is never really that challenging, but what it can do (if you allow or encourage it to) is set up some really interesting scenarios.

[Pode]

You may well have a point, but IMO it does not undermine the fundamental reason for this test. We are attempting to test if the AI acts differently in the two cases, where the player is triggering AI in exactly the same way, with the only variable being the saving/loading of the game in one case.

The goal seems to be to confirm statistically that there is in fact a difference, and only then evaluate whether that difference has a gameplay impact, correct?

[roguebolo]

Well, something must trigger the AI to be aggressive again. Possibly activity on the part of the player?

To be fair, I must admit that I never "sit out sieges". I always rally, usually on the turn after being besieged and never later than the turn after that.

I should mention that in my current campaign, which I'm playing as the Julii because I reinstalled RTW and Patch 1.2 from scratch, I still have the save games from just about every turn in the game. I think I will go back and take a look at them and turn off FOW and see what's happening with the OTHER factions -- the stuff I can't see.

*edit*

Actually, I just thought of another load game issue which I've always thought was unrelated but might not be. You know how random events can happen when you select End Turn? Well, if you don't like the outcome of those events, you can do a load game. If you just hit End Turn again, you usually get the same random events. But if you make ANY change -- say, move a noncrucial diplomat or something -- it will re-roll the dice for all of those random events, and you will get a completely different set of results.

Perhaps this is related, perhaps it is not. But the change definitely occurs only if there is player activity AFTER the load game and BEFORE the End Turn.

[roguebolo]

Looking back through my save games, I have found that I saved each and every one of my first twenty moves immediately prior to hitting "End Turn", so I can track each faction's activities for the first 20 turns. However, FOW doesn't seem to want to be disabled if it was originally enabled.

By turn 20, the majority of the nations were still in control of only their starting territories and had not engaged in warfare with other factions, with the following exceptions:

1.) The Seleucids and Egypt were at war, Egypt gaining a territory on turn 11 and losing a territory on turn 18; the Seleucids gained a territory on turn 18.
2.) Parthia and I (Julii) both gained a territory on turn 2.
3.) The Brutii gained a territory on turn 4 and another on turn 14.
4.) Pontus gained a territory on turn 14 and another on turn 16.
5.) The Greek Cities were at war with Macedon and Pontus.
6.) The Seleucids were at war with Egypt and Armenia.
7.) Gaul and I were at war.
8.) Brittainia and Germania were at war.
9.) Scythia and Parthia were at war.

Now, that's not exactly inert inactivity, and there were a considerable number of load games during those first twenty moves. I think the fact that I can upload each and every save game to whomever wants to look at them, as well as numerous variations I tried throughout the course of the campaign, is proof of that.

So far, I managed to find only one save game where I was under siege, in Thapsus, later in the game. There was another Numidian army approaching Carthage, obviously intent on laying siege, at the same time. As usual, I sallied when I was actually playing the game. However, if I load the save game and simply hit "End Turn" without sallying forth, the Numidians do indeed abandon the siege -- to block the bridge between Thapsus and Carthage, the only way I can get reinforcements there, while the other army layed siege on Carthage. Note that the second army did not forget that it's plans were to lay siege, and that the first army did move to a logical strategic point! (Except for the fact that I had a trireme in dock at Thapsus.) However, here's the bummer:

If you do a save/load at that point and hit "End Turn", Numidia also abandons the siege of Carthage and both armies walk away. If you just hit "End Turn" without the save/load, then they maintain the siege. So there's apparently some validity to this observation and it is some type of player activity which is defeating it. I will look at some additional save games and see if I can identify it. Anyone else who wants to look at them is welcome.