Investigation of AI reassessment upon reload

[Epistolary Richard]

I don't quite understand what you're suggesting. What would people do differently to try and work around the issue? Just play for longer?

[sunsmountain]

Don't use the quick save at all! Use the normal saves, and it should be okay! Are you guys planning on testing that as well?

ie, 15 turns of doing nothing, versus
15 turns of saving and reloading using normal saves?

Maybe quicksave was MEANT not to save the army status (standing or laying siege) when on the campaign map, therefore you lose the siege, but you gain the benefit of quicksaving and loading.

[roguebolo]

Um, I hate to say this, but the proposed test does not seem to impose any controls which eliminate other possibilities.

For instance, is it possible that as part of game theory that the AI factions are approximating the aggressiveness exhibited by the human player? This is fairly commonplace in computer games. For one thing, it allows players of all different types to stand a chance at winning the game. For another, it offers gameplay which is compatible with the style of the human player. If the human player merely "turtles in", why shouldn't the AI do the same? There is no percieved threat as the human player is not yet being aggressive.

Second, the AI might use a "delta factor" to determine it's reassessment of moves -- in other words, the degree of change from the previous move prompts the AI to make decisions regarding the current move. Maintaining the status quo, as the human player seems to be doing, is not actually a losing scenario. With a delta of zero, the AI factions have an expansion factor of nearly zero.

The fact of the matter is, I have performed the proposed test and came up with identical results -- none of the territories change hands. However, in a REAL game, in which I actually do save and load every single turn -- usually multiple times per turn -- and which I play aggressively, each AI faction actually does seek to expand its empire.

So I don't think this test is a valid proposal. It needs more thought.

I'm a seasoned programmer myself and I've read through some threads by other programmers theorizing on the nature of this problem and one of them was pretty knowledgable. He stated that in order for the AI to remember its long term goals that its planned actions would need to be "serialized" to disk (meaning, stored in the save file along with the positions of units, etc.) which is a nontrivial programming issue because the information is stored in memory using objects that are linked together by pointers. He's probably absolutely correct about the data representation, but the truth is that serialization of such objects is actually not a complex task. I have done it numerous times with numerous software products -- and new languages like C++ make it a little easier since it's sort of a built-in feature of the language. It's actually a rather mundane and simple task when compared with some of the more complex programming that is tackled in a computer game, as another programmer pointed out to him. So I don't think the explanation that the programmers thought it was too difficult or too much work to save information in the save file is a good one. It would only seem the obvious thing to do so that a save game starts where it left off and merely requires a little grunt work.

One issue which even most of the programmers were a little confused about seemed to be the idea that the game looks ahead a given number of turns to make long range plans -- some of them were stating possibly as much as 20 turns. Have you ever played chess? One of the measures of a good chess player is how many half-moves he can look forward into the game. Five is considered good and I believe about seven or eight is the record for a human being. Chess programs do the same thing. So looking ahead 20 full turns is a little unrealistic on a map this size and with this many pieces and factions ("players"). The AI factions might have general goals but they cannot plan ahead for specific movements for too many turns because they need to react to the human player and the other AI factions. So it's not an issue of serializing detailed, long-term movement plans for a projected 20 moves; simply a matter of storing generalized objectives in the save file.

Taking all of this into consideration, if the AI is reevaluatiing its position as CA claims, then it must have elected a NEW set of campaign objectives when it relieves its sieges. These new objectives must take into consideration the fact the besieging faction is at war with the besieged faction and that trade is no longer possible with them. It must be able to find a more desirable alternative in light of that problematic issue. Does it attempt to negotiate a ceasefire and/or trade rights in the same turn? It would be interesting to analyze a move in which the AI has abandoned it's previous objectives and attempt to determine which objectives it has identified as more desirable.

By the same token, if the AI is reevaluating its military objectives it must also reevaluate its economic objectives, since the two go hand in hand. Is it canceling or changing building and economic development to match the military changes?

Also, does anyone know where I can read CA's official response regarding this issue and other bug reports?

[slackker]

rouge all CA currently stands in this issue is found here: http://p223.ezboard.com/fshoguntotal...ID=24377.topic

[Colt45]

Currently, the tests Ive seen are too broad, given the large amount of factions in the game. I propose that someone mod a game to be setup with 2 factions, plus senate. Use an isolated area and put both factions there, i.e. Britain. Give AI large army and yourself a minimal garrison. Now let him siege you and perform the save/load system. Make sure there are no rebels in his/your area for accuracy. Report results.

I feel, given the CA description, that you need to limit the variables much more than just starting a campaign with all factions on board. I feel this would definitely give a much better example.

[ShellShock]

As well as testing the effect of reloading every turn, I suggest we also test at other frequencies of reload, e.g., every 2, 4, 8, 16 turns (I know this is a lot of testing!). I know some people have done this elsewhere and they show evidence that the expansion of the AI is directly proportional to the frequency of reloads - this should produce a very convincing graph (number of provinces changing hands by rate of reloads).

I can imagine that CA may respond to a test of reloading every turn as being unrealistic (although I for one often only play one turn at a sitting). We can counter this by testing at other frequencies of reload.

Also I suggest we use the faction rankings graph (rather than removing FOW) to assess the impact on the AI. To simply things we could use the Top 5 ranking on e.g., Territory - that is, make a note of the number of territories each of the top 5 factions has. The beauty of the graph is it shows historical data too, so at the end of the test you can see how the AI expanded for each turn.

[roguebolo]

I think it is a confirmed fact that loading a saved game causes RTW to regenerate (or recalculate) random events and often changes the behavior of the AI, especially regarding sieges, and that it is supported by empirical and repeatable data.

My own tests were identical to the one that was proposed here, except that I started at various midgame points wherein alliances, declarations of war, and faction targets and strategies were clearly established to see what kind of impact this actually had on the game. I would have reverted to an earlier patch if it had turned out to be significant. However, I'm pretty convinced that this is not the case.

The problem with testing the phenomenon when no moves have been made by the player at the opening of the game is that it is testing an extrema which generally only happens once in a game -- at the first move. That is the point in the game where no alliances have been established, no warfare, no movement, and no building. Many computer algorithms exhibit awkward behavior with various extrema -- for instance, the quicksort (which is the fastest sorting algorithm on the average for most data types) exhibits its worst-case (slowest) behavior with data which is already sorted. That's not really a concern when the programmer expects the program to be working with unsorted data. Extrema can be very useful in determining actual bugs in a program such as integer overflow or divide-by-zero errors. But what's being analyzed here, the AI, is a complex network of heurisms and algorithms. The truth be told, the AI is probably playing a better game than the human player during this test, because it's undoubtedly building an economic infrastructure for each faction and moving its diplomats. It might be possible to force some military stagnation at the most extreme moment in the game through save/loads and nonactivity, but under more realistic conditions I found that it was not as easy to accomplish. Even in the first 20 moves, with multiple save/loads per turn I was unable to recreate that kind of stagnation. In midgame the proposition is even more difficult.

Knowledge of the fact that this reassessment logic is tied to the load game feature is surely something that I can exploit if I wish, but I am summarily unconcerned with any other ramifications of gameplay.

Finally, I'm not sure that it's really such a good idea to have the objective of just a few large AI factions in endgame, as was always the case with MTW. In the early stages of MTW, the AI was geared to force the player into situations modeled after historical realities. The historical reality is that Rome conquered each of these rivals individually, and ultimately become entangled in civil war.