AI analysis: The first 100 turns

[MadmanAtW]

So I see that a couple of builds there are "City Defense Sword". Is that a good idea? Swords are not a unit that one usually thinks of as defenders.

 

[jdog5000]

Yeah, defensive swords is one thing the AI does which I never would ... in terms of strength, they seem like a decent choice in the archer era which is probably why the AI chooses them. They're an okay defender against other swords or horsemen but not the best against either. The AI would be better off IMO with a combo of axemen and spears though, they're cheaper and cover all the bases in the earlier game.

 

[jdog5000]

Turns 90-99

The last interval of this test.

De Gaulle

Techs: T90 - finishes Monotheism, selects Monarchy; T91 - chooses Metal Casting as free Oracle tech;
Paris builds: T91 - finishes Oracle (2 of 7 pop rush), selects Forge with production 1; T93 - finishes Forge (1 of 6 pop rush), selects state religion Missionary; T94 - finishes Missionary, selects Worker by method 3; T98 - finishes Worker, selects city defense Archer; T99 - finishes city defense Archer, selects Settler by method 1
Orleans builds: T91 finishes explore Work Boat, selects Forge with production 2; T97 - finishes Forge (1 of 4 pop rush), selects counter Chariot; T99 - finishes counter Chariot, selects Settler by method 1
Lyons builds: T91 - selects Monument; T99 - finishes Monument, selects Granary
Other: Found Lyons on turn 91, beating Monte by a couple turns to the spot between their empires. Not sure when it happened exactly, but Paris picks up Monte's state religion.




Montezuma

Techs selected: T92 - finishes Mathematics, continues on to Construction
Tenochtitlan builds: T93 - finishes Library, selects Walls; T96 - finishes Walls, selects city defense Archer; T97 - finishes city defense Archer, selects attack Jaguar; T99 - finishes attack Jaguar, selects city defense Axeman
Teotihuacan builds: T90 - finishes Worker, selects Temple; T95 - finishes Temple, selects Walls
Tlatelolco builds: T90 - finishes Settler, selects Barracks; T95 - finishes Barracks, selects city defense Jaguar
Texcoco builds: T97 - finishes Granary (1 of 2 pop rush), selects Lighthouse
Other: After Lyons is founded, Monte redirects his settler farther North. The settler appears to disappear on turn 94, perhaps the victim of a Barbarian attack. Monte then dispatches the settler he complete on turn 90 in Tlatelolco towards that Northern site on turn 99 after a spare Archer comes over to escort.

Notice how Monte actually picked out and bee-lined for a tech (Construction) that he first had to research two prereqs for. Interesting. Also, Monte builds several more advanced buildings during this interval, the first to both Libraries, Temples, and Walls.




Zara Yaqob

Techs selected: T98 - finishes Monarchy, selects Writing
Aksum builds: T90 - finishes Settler (2 of 6 pop rush), selects city defense Axeman; T91 - finishes city defense Axeman, selects attack Sword; T95 - finishes attack sword, selects city counter Spear with floating defender 1; T97 - finishes city counter Spear, selects Settler with method 1
Gondar builds: T92 - finishes Missionary, selects Settler by method 1
Lalibela builds: T90 - finishes Lighthouse (1 of 2 pop rush), selects Granary
Addis Ababa builds: T95 - selects city defense Archer
Other: On turn 93, Zara sends a settler to the west. Addis Ababa is founded on turn 95.




Standings after 100 turns:

Montezuma

Power: 134

Total score: 993
Pop score: 504 (23 total pop)
Tech score: 137
Land score: 336
Wonder: 16


Zara Yaqob

Power: 90

Total score: 830
Pop score: 328 (15 total pop)
Tech score: 113
Land score: 357
Wonder: 32


De Gaulle

Power: 91

Total score: 612 (10 total pop)
Pop score: 219
Tech score: 119
Land score: 226
Wonder: 48

 

[jdog5000]

So, there we have it ... 100 turns under the hood of the AI. How'd they do?

What surprised you or popped out at you?

Anything you thought they did well?

How would you have played these starts differently?

 

[r_rolo1]

First thing that came to mind reading the last set :city defense jaguar City defense sword is already bad, as you said above, city defense jaguar is just stupid unless you don't have metals, which is not the case.

 

[MadmanAtW]

I'm not sure anything's jumped to mind that hasn't already been said, but I wanted to say that I loved this view "under the hood". Fascinating stuff.

 

[Afforess]

I'm not sure anything's jumped to mind that hasn't already been said, but I wanted to say that I loved this view "under the hood". Fascinating stuff.

Yes, I agree. It's been very revealing. I think the notes here could provide months worth of better AI changes...

 

[jdog5000]

I think the notes here could provide months worth of better AI changes...

No kidding!

 

[jdog5000]

Here's another idea ... ask not what the AI can do for you, but what you can do for the AI:

If any of you are interested, I think it would be really fantastic to compare how you would play one of these starts as compared to the AI. First 100 turns only.

How you want to go about it is totally up to you. You could try and wipe your mind clean of what you saw in this thread and play one of the starts in your own way. Or, you could try you hand at playing how the AI did broadly and seeing how much farther you get ... so, you might going for buildings and wonders as De Gaulle, do religion and tech like Monte, or try settler first like Zara. Or, you could try playing how you thought the AI should have, like rushing early with Monte or something.

I've started a new thread for these so that we keep the discussion on human vs AI decisions separate.

 

[Zeiter]

A few ideas I got after reading this intriguing AI walkthrough:
1. Make the AI value a bit more highly the prospect of continuing on with a build and finishing it if the AI has already sunk hammers into the item. AI should still have the flexibility to abandon production items when really needed, but try to give the AI some incentive to be less capricious about changing production just because the city grew a pop, especially if a build already has like half of the hammers invested. Don't waste those hammers, if at all possible! The AI should value staying with a build more and more vs. switching mid-stream to other options the more %hammers the AI has invested in a build.

2. Make sure, if an AI has an idle settler due to lack of escort, that building an escort unit receives high priority.

3. Every time the AI thinks about producing a missionary, the AI should look at how many X turns it will take, at the present rate, to build the missionary. Then the AI should estimate the number of cities without that religion (domestic or foreign) that the missionary will be able to access in X turns (by taking current # of accessible cities without religion + how many new settler units, at current production rates, will be produced in X turns, and assuming no religion auto-spread in the meantime. Thus, for example, if an AI wants to build a missionary in city Y, it finds that it could theoretically do so, at present rates, in 5 turns. The AI sees none of its cities currently without that religion. However, the AI finds another city, city Z, that is producing a settler to be finished, at current rates, in 3 turns. So then the expected number of accessible cities at time X (5 turns) is 1. So if the AI doesn't already have at least 1 missionary on hand, then it is free to build another. Otherwise, skip to next build option).

 

[TheMeInTeam]

1. Make the AI value a bit more highly the prospect of continuing on with a build and finishing it if the AI has already sunk hammers into the item. AI should still have the flexibility to abandon production items when really needed, but try to give the AI some incentive to be less capricious about changing production just because the city grew a pop, especially if a build already has like half of the hammers invested. Don't waste those hammers, if at all possible! The AI should value staying with a build more and more vs. switching mid-stream to other options the more %hammers the AI has invested in a build.

Sunk costs are not relevant to decision making. If switching is better, it should switch. If not, it shouldn't. There isn't a need for that kind of weighting to deliberately avoid sensible economic criteria in favor of sunk costs.

The other suggestions sound good.

 

[Panopticon]

Sunk costs are not relevant to decision making. If switching is better, it should switch. If not, it shouldn't. There isn't a need for that kind of weighting to deliberately avoid sensible economic criteria in favor of sunk costs.

The other suggestions sound good.

It depends whether the AI takes the "true remaining cost" of its choice into account. If choices are based solely on an attraction value of some kind, then it may make sense to account for sunk costs, simply to gauge the true return per hammer. Sunk costs don't affect the benefit part of the trade-off between two choices, but the AI should be willing to continue with a build order with lower absolute return if it offers an acceptable return per hammer.

The caveat being that I don't know how these choices are made as well as BBAI people.

 

[MadmanAtW]

Sunk costs are not relevant to decision making. If switching is better, it should switch. If not, it shouldn't. There isn't a need for that kind of weighting to deliberately avoid sensible economic criteria in favor of sunk costs.

The other suggestions sound good.

The only thing it should try to do is not let sunk hammers decay if it is going to want that building/unit later. Which is probably not really worth coding in.

Is the amount of time it will take to build something a factor in how likely it is to be selected? Is a Forge more valuable to an Industrious civ than a non-Industrious? A 3/4 built Granary should be worth more in the valuation than an unstarted Granary. That's different than worrying about sunk costs, it's taking into account that, right now, you get the benefit for the building sooner, for whatever reason.

 

[TheMeInTeam]

It depends whether the AI takes the "true remaining cost" of its choice into account. If choices are based solely on an attraction value of some kind, then it may make sense to account for sunk costs, simply to gauge the true return per hammer. Sunk costs don't affect the benefit part of the trade-off between two choices, but the AI should be willing to continue with a build order with lower absolute return if it offers an acceptable return per hammer.

The caveat being that I don't know how these choices are made as well as BBAI people.

No, you simply never use sunk costs. The RELEVANT costs involved in this case are the remaining costs of the invested build and possibly hammer decay, with the benefits of the current investment compared against those of switching.

Any hammers already invested *need* to be disregarded in this process. The only relevant hammers toward the item being built are the ones that are not yet invested and thus can vary between decisions.

 

[Elandal]

Sunk costs are psychological issue in most cases, and in real world the psychological issues unfortunatelly lead to suboptimal decisions quite often. Hammer decay may be hard to value properly, so remaining cost is the relevant issue.

 

[Panopticon]

No, you simply never use sunk costs. The RELEVANT costs involved in this case are the remaining costs of the invested build and possibly hammer decay, with the benefits of the current investment compared against those of switching.

Any hammers already invested *need* to be disregarded in this process. The only relevant hammers toward the item being built are the ones that are not yet invested and thus can vary between decisions.

Yes, but do we know whether the AI takes remaining hammer cost into account? If it doesn't, then it ought to do so, and that's when sunk costs become relevant, because they are isomorphic to remaining costs on a proportionate level. (No need to shout; I have spent much of several degrees' worth of study investigating cost-benefit analyses.)

 

[TheMeInTeam]

Yes, but do we know whether the AI takes remaining hammer cost into account? If it doesn't, then it ought to do so, and that's when sunk costs become relevant, because they are isomorphic to remaining costs on a proportionate level. (No need to shout; I have spent much of several degrees' worth of study investigating cost-benefit analyses.)

Sure, if you want to use complicated terminology to point out that they're only relevant to the extent they decay.

But it's probably simpler/equally effective to just consider decay as a switching cost, avoiding the consideration of any invariable cost and lowering the chance of confusion. Certainly if the AI considers this scenario it would be easier to figure it as I describe?

 

[Panopticon]

I think we are agreeing but using different terms, so this isn't a particularly productive discussion.

 

[MartinHarper]

Missionaries are not wasted hammers.

In general, no. In the specific case of Montezuma on turns 70-80, yes. He has a good chance of getting free religion spread from trade, the border pops aren't important, he's making monuments anyway, and he isn't in a religious civic. It would be better to build those missionaries later. Meanwhile, a successful military offensive at this stage will effectively win Monty the game.

 

[Janov]

I agree, the hammers already spent are a thing of the past. However, the hammers remaining to complete whatever is being built should be factored in. Let´s say you need two turns to complete a really expensive wonder - but are under attack from a fairly weak civilization. Continue the wonder, THEN produce defenders, or do so right away?: The appearant value should be "benefit" divided by "hammers to complete". So even a fairly small "benefit" could reach a high value, if the "hammers to complete" is small.
So it is not a trivial decision for the AI, and you will find even humans coming to different conclusions. Switch to Archer, and let someone else get Stonehenge? Maybe I have enough defenders after all? If I get a lucky roll, they will hold long enough... but, on the other hand, if I loose that city? Play safe, or gamble? Finish that missionary (I WILL need one eventually later), or put everything I have into that Axemen-rush?

Not a quick fix for jdog to make here, I think.

Janov