Transport Tycoon Forums

So am I correct in assuming the current roster of AIs is as follows?

Maintained Miscellaneous
CivilAI
TownCars

Bug Free, Unmaintained Miscellaneous
Idle
IdleMore

Maintained Competition
AIAI - Trucks, Busses, Planes, Trains
AroAI - Trucks, Busses
BorkAI - Planes, Helicopters, Trams, Busses, Trucks
ChooChoo - Trains
CluelessPlus - Busses, Trucks, Planes, Helicopters
DictatorAI - Busses, Trucks, Planes, Helicopters
NoNoCAB - Trains, Trucks, Busses, Planes, Ships
OtviAI - Busses, Trains, Trucks, Trams, Ships
RoadRunner - Busses, Trucks, Trams
SimpleAI - Busses, Trains, Trucks, Planes
SynTrans - Busses, Trucks, Planes
Terron AI - Busses, Trucks, Aircraft, Ships
trAIns - Trains
WormAI - Trains, Planes

Bug Free, Unmaintained Competition
CPU - Busses
gelignAIte - Busses, Trucks
Rythorn Airline AI - Planes, Helicopters
WrightAI - Planes, Helicopters

I haven't checked all AI's but at least SynTrans and OtviAI have several open issues where the author hasn't responded for several years. I don't think they are actively maintained.

I have seen Rythorn crash and it has some other problems so it certainly is not bugfree.

Although Admiral AI has also some open issues I think crashes are pretty rare. Maybe not bugfree but certainly low on bugs and playable in my opinion.

My own WormAI and NoNoCab are maintained although they have some known issues but no known crashes. I'm hoping to get some time to look at that later this year.

Bumping...I noted a recent update in SimpleAI.

I have a lot of spare time (yay, disablementy things) so I can try to come up with an AI, or figure out how they work and add to a free licence one that allows derivitive works.
I didn't know people actively used AIs, as all the ones I have used make very predictable moves, mixed with sometimes illogical or broken routes, I at least stopped using them for that reason.

Baldy's Boss wrote:Bumping...I noted a recent update in SimpleAI.

Did you also update the wiki page?


@Redirect Left: Fixing or improving an existing AI would be more useful I think. We have lots of AIs, but very few maintainers. That also gives you a head start, as you have to do less ground-work.

Alberth wrote:Did you also update the wiki page?

If you mean this one, then yes, I've already done it with the AI update

No big flashy new features though, it was just a minor fix related to airplanes.

Alberth wrote: @Redirect Left: Fixing or improving an existing AI would be more useful I think. We have lots of AIs, but very few maintainers. That also gives you a head start, as you have to do less ground-work.

Well, i'll have a look around, and if I find one that users seem to particularly enjoy that I can add to or resolve a glitch with (it'll take a while to learn AI code anyway) I'll see if it allows derivitive works / outside corrections, improvements & additions.

I note there was an update of ChooChoo since the last post or update here.

Bumping topic to the front,we need an updated survey...I note entirely new AIs have come along,like RailwAI.

I just did some testing with the current competitive AIs and these were my results (Version 1.9)

I tested on the 3 default maps (Not toyland) with vanilla assets only
Start year was either 1950 or 1970
I kept games going for around 5 - 7 years
Max loan was $500,000


Excellent- Both consistently performed very well
NoCAB
AIAI

Good - Sometimes performed well and sometimes very well (Unfortunately 3 of them crashed occasionally)
Wright AI - depended on the presence of large towns/cities
ChooChoo
Grinchios - crashed
Fast Transport- crashed
TracAI - Crashed

Medium - Usually performed pretty well. Better than those below
AdmiralAI
Trans

Average - turned profits but did not perform to the levels of those above
Terron
trAIns
WormAI


Low - stayed in business or went out of business
RailwAI
Clueless
Pathzilla
OTVI -crashed
Denver & Rio Grande -crashed
NoNoCAB


Failed - often went out of business
Rythorn -crased quickly
CivilAI

I'm sorry to hear CivilAI performed so poorly for you. I'm currently working on a major update for OpenTTD 1.9, so if you have any example savegames in which the AI is struggling I'd be interested to see them.

I would be interested in savegames too.
For WormAI my guess would be that it is usually not the fastest starter and will perform better on longer runs.
For NoNoCAB it's more surprising. In general it should perform similar to NoCAB since it is based on that.

What limits on vehicles did you have, ai difficulty setting, size of map, terrain and amount of water are settings that might influence performance of ai's.

Edit:
Is Grinchios a new AI. I don't see it in the online content.

Redirect Left wrote:

Alberth wrote: @Redirect Left: Fixing or improving an existing AI would be more useful I think. We have lots of AIs, but very few maintainers. That also gives you a head start, as you have to do less ground-work.

Well, i'll have a look around, and if I find one that users seem to particularly enjoy that I can add to or resolve a glitch with (it'll take a while to learn AI code anyway) I'll see if it allows derivitive works / outside corrections, improvements & additions.

Unfortunately I had to quickly give up on this idea, as i found the way AIs are coded to be very unituitive and hard to figure out exactly.
I hope in the future, the method of implementing AI is improved to be a lot easier to do. Building a fully working and decent AI that builds proper networks and is very profitable seems to be more effort than the pleasure you'd get out of it at the end, at least from what I saw.

Redirect Left wrote:I hope in the future, the method of implementing AI is improved to be a lot easier to do.

It's a C-like scripting language - it's hard to imagine how it could be any simpler. Granted, the way 'real' programmers have structured their AI is often impenetrable to naive rubes like you and me, but that's not the language's fault.

Redirect Left wrote:

Redirect Left wrote:

Alberth wrote: @Redirect Left: Fixing or improving an existing AI would be more useful I think. We have lots of AIs, but very few maintainers. That also gives you a head start, as you have to do less ground-work.

Well, i'll have a look around, and if I find one that users seem to particularly enjoy that I can add to or resolve a glitch with (it'll take a while to learn AI code anyway) I'll see if it allows derivitive works / outside corrections, improvements & additions.

Unfortunately I had to quickly give up on this idea, as i found the way AIs are coded to be very unituitive and hard to figure out exactly.

Yes, they are, but at the same time they are not.
The main reason why you find the code unintuitive and hard to read is because you didn't write it. Code follows the logic of its author, and everybody has a slightly different logic. It takes time and effort to figure out what logic the code has and to make that your own. In my experience, if you stick with studying it, you'll start to see patterns at some point, and then slowly code starts to make sense.

Of course this also goes in the other direction. Code that you or I write looks unintuitive and hard to understand to most others. It's not bad code, it's just how it works.

I hope in the future, the method of implementing AI is improved to be a lot easier to do. Building a fully working and decent AI that builds proper networks and is very profitable seems to be more effort than the pleasure you'd get out of it at the end, at least from what I saw.

I hope too that it gets easier, but I don't expect that to happen in my lifetime tbh unless you throw a lot of CPU power at it (ie deep learning and such). Writing a good AI is just a hard problem.
Ever tried to find a computation that understands when a network is "proper and very profitable"?

For most AI authors, I think the end-result is less relevant, but the journey or path to the end is. The pleasure of learning new things, and finding how to fix that problem in a nice way.

Alberth wrote:[..] snipped excess quotes [..]

I think it might be i was trying to look at the ChooChoo AI because it (in my opinion) builds the best networks similar to natural players. So it was quite a lot of code, that is no doubt perfectly organised to the author.

Unfortunately based on that, it takes a lot of effort to get a network builder that is competent. So huge props to the author of that. A few AIs don't specifically note the licence they're released underfor editing / improving and releasing either so i can learn that way. I tend to learn code by seeing and editing existing pieces of work, rather than from scratch.

Wormnest wrote:I would be interested in savegames too.
For WormAI my guess would be that it is usually not the fastest starter and will perform better on longer runs.
For NoNoCAB it's more surprising. In general it should perform similar to NoCAB since it is based on that.

What limits on vehicles did you have, ai difficulty setting, size of map, terrain and amount of water are settings that might influence performance of ai's.

Edit:
Is Grinchios a new AI. I don't see it in the online content.

PikkaBird wrote:I'm sorry to hear CivilAI performed so poorly for you. I'm currently working on a major update for OpenTTD 1.9, so if you have any example savegames in which the AI is struggling I'd be interested to see them.

Ah, sorry guys, I didn't save the games
I'll remember to do so later

Settings:
Usually the Temperate Map (But also desert and arctic)
40 towns
Low industries
500,000 max loan
2% sea
flat terrain
medium rivers
1024x1024 size map
No custom assets, just vanilla vehicles and carriages


For CivilAI, it seemed that it was a little reluctant to start routes as it was always searching for the best one. Unfortunately, since I usually put the AIs up against 5-6 others, they were crowded out by more aggressive AI
I was surprised about NoNoCAB too, that the original version seemed to perform much better in which routes to go for
WormAI was actually pretty good, just not as good as the ones listed above it
I think grinchios is listed in the AI page under mungo (But I cant remember)

Sorry for necroing this thread. Can anyone confirm which AI can survive thorough 50 years without crashing? I tried out 9 AIs and 2 of them crashes randomly:
--- No crashes so far ---
- LudiAI
- NoNoCab
- KrakenAI
- SimpleAI
- CivilAI

--- Random Crashes ---
Very good AI. Unfortunately, the crashes seems so random that I couldn't reproduce it.
- TransAI
- RailwAI

--- Very suspicious ---
- AdmiralAI (Haven't crashed in my game. However the forum indicates that this AI has been abandoned for long time and people just reporting that this AI would after 20 years)

--- No crash but Uninstalled anyway ---
- TrAIns (no save/load support resulting in eternal AFK from company controlled trAIns)

I didn't try the rest since the last post of the AI forum was just bug report without any solution from the author. I also want to say big thanks to all of the authors that had created this AI in their spare time and I hope we can get better AI in the future. Thanks and stay safe everyone!

Nobody can give you a warranty that an AI would never crash, not even its author. The more complicated a program is, the harder it is to predict its behaviour in every possible situation. For any program except the simplest ones, this becomes virtually impossible.

In professional software development, this problem is partially mitigated by extensive testing, both automated and manual, but I don't know to what extent this approach can be applied to game AIs.

If you need a rock stable AI, the best thing you can do is pick one of those that are still maintained, and report every crash you experience. Hopefully, the maintainer would find the time to fix the problems you find.

Too bad most of them left their AI unfixed since I play with AI all the time and I love AI very much. I wish I could make my own AI but I hate how complicated C is and with these unfixed AI have enough to tell me that it's very hard to create those.