Investigation of AI reassessment upon reload

[roguebolo]

I did not actually test twenty turns without the save/loads from Turn 1 because I have faith that I will get the same results that have been posted by others; the reason why is that I have a glimmering of the kinds of algorithms that were used and why and how the load game behavior is resetting the AI's behavior. It essentially feels free to select a new opening move. The available options are probably somewhat similar to a software random-number generator, which does not truly generate a random sequence of numbers but eventually will cycle through the sequence and start at the beginning again. While the test results are interesting, they seem to be specific to a given set of extrema.

The control, or the factor that was eliminated in my tests, was player inactivity. A review of the first 20 moves of my campaign -- each and every one of which was saved, because as I said I do multiple saves and loads each turn -- shows clearly that the AI factions are also seeking to expand. I listed the exact details of their expansion and other interactions with each other and with the human player (except alliances) in a previous post.

To test further, I introduced player inactivity at midgame save points, wherein alliances, declarations of war, sieges and other policies had already been initiated, and tried hitting "End Year" repeatedly both with and without save/loads. I posted those results as well, but I will summarize here. Although the AI did have a general tendency to relieve sieges after a save/load, I found cases both where it did so and where it did not; and in those cases where it did relieve a siege, it did not abandon its objective but merely changed the way in which it manifest itself as the aggressor. Units in route to a siege consistently continued toward their target, even with save/loads; units that relieved a siege because of a load game consistently continued to threaten the territory and sometimes blocked access to other besieging units for reinforcements; in some cases they abandoned their siege to add support to another siege in which they were undermanned. I found no cases in which they yielded an inferior strategic position as the result of relieving a siege -- and quite honestly, they were more annoying when they were relieving and reinstating the sieges, almost as if they were taunting me!

I have some really interesting save games that will allow you to observe the phenomenon yourself. It's actually kinda funny, watching the British in the north and the Numidians in the south alternately engaging and disengaging sieges each turn, which gets triggered by the save/load, but it also becomes pretty apparent that they are no less of a threat to your Empire than if they maintained the sieges as they do when you hit "End Year" without a save/load. That's when I realized that if I were willing to endure a siege for just one turn I could use a save/load to exploit this AI behavior to meet them in the field rather than sallying forth from the castle. However, that's probably not a very useful exploit since they will probably end up with the terrain advantage and I will lose the castle which can protect my flanks as I approach them. My guess is that this is one of the factors in the AI's assessment that relieving the siege is equally acceptable to maintiaining it; they can choose which square (or which realtime map) they wish to defend, giving them the advantage of position if I decide to attack them. That's always been an advantage of the AI, that it has a database of which realtime battle map will be used for any given square of the campaign map.

The reason I chose Turn 19 for my other test was because it was still early enough in the game that there had been some interaction between the different factions, but not extensively. I wanted to find out precisely how close to a position of extrema you needed to be in order to observe more stagnant behavior on the part of the AI as the result of player inactivity and load games. I expected results fairly similar to starting the test from Turn 1, but perhaps not quite as pronounced. However, my testing was rudely interrupted after just a few turns when the Brutii laid siege to the rebel town of Segistica. After the first save/load they relieved the siege, then after another save/load they reinstated it and after subsequent save/loads they maintained the seige and occupied the territory. They do that consistently, so in the reassessment logic there must be no "equally viable" alternative to maintaining the siege. The one interesting thing that I noticed is that when the AI relieved the siege after the first save/load, then reinstated it, it had no effect on the attrition of the settlement! The settlement could only hold out during the siege for 3 turns rather than 4, just as if the siege had never been relieved!

For players that play only on turn per sitting, I suppose this behavior could have a small advantage in that sieges on their settlements will be audibly announced every other turn.

Anyway, as you've probably guessed by now, I'm a programmer myself. After reading the response from "The Shogun" at http://p223.ezboard.com/fshoguntotal...ID=24377.topic it seemed to me as if he had brought up the topic to his programmers, and as if they fully understand the phenomenon and explained it to him, but as if he did a poor job of explaining how a load game causes the AI to reset the list of equally viable options -- selecting the "seed value", to describe things in terms of a random number generator. He also did not explain the fact that although this behavior can produce some awkward gameplay at the moment of greatest extrema (which happens to be the first turn of the game in the presence of player inactivity), those side-effects diminish as the game moves forward -- and that it can, in fact, enhance the gameplay experience because the computer cannot be relied upon to make the same move after a load game as it did when you simply pressed "End Turn", thereby making it less predictable.

My guess is that the developers showed him some concrete examples, like the ones in the save games from my campaign that I've experimented with. That's why I've offered to send them to anyone who wishes to review them. They were not specifically constructed to create a point; they are just save points from my campaign that I thought would be good points to test the effects of this phenomen during midgame -- specifically, those points at which sieges were in effect. Combined with a little explanation, they are sufficient evidence of the viability of strategic reassessment after a load game. Unfortunately, I think something was lost in the translation from developer to customer relations to end user. Have you ever played the game "Secret"?