Any "casual" AI available?

[sucinum]

This looks promising so far. The AIs (7!) were quite strong and built many different lines throughout the map. I didn't interfere a single time.

Train 4 and 5 were created after the AI had established there, but i wouldn't have complained the other way round. That bus connection next to those trains is something you should have a look on. The track is designed faulty: too curvy, unnecessary ascends/descends and generally too long for busses, so altogether that causes a lot of breakdowns. Ofc it's a very huge advantage over the vanilla AI and good enough. But still room to improve.
I tend to cap bus/truck connections at about 50-60 tiles. The short-range-trains i construct are much more efficient, even though construction is more expensive (especially since i'm very anal about straight lines and waste loads of cash even in the beginning).
There is also a heavy jam at the nondwood station (sp?).

Loading the game took ages and also the acceleration button had less and less effect throughout the game. I hope this doesn't get worse in the late game. Got a Pentium 4 (2 GB RAM), which should really suffice. And then there could occur the other player who prefers 2048² with 15 AIs...

I still managed to get crashes for the AI, but i guess that's a doable bugfix (div by zero). I added savegames to check.

PS: I am still experimenting with waypoints and other stuff to be able to use multiple trains on a track. That's something I didn't do before meeting the OTTD comm, so still reading and learning. But seems to be working. As said, i'm a very casual gamer, but always willing to improve. But even with obvious flaws in design i beat the AIs (not complaining about that at all, just saying!).

[Morloth]

This looks promising so far. The AIs (7!) were quite strong and built many different lines throughout the map. I didn't interfere a single time.

Train 4 and 5 were created after the AI had established there, but i wouldn't have complained the other way round. That bus connection next to those trains is something you should have a look on. The track is designed faulty: too curvy, unnecessary ascends/descends and generally too long for busses, so altogether that causes a lot of breakdowns. Ofc it's a very huge advantage over the vanilla AI and good enough. But still room to improve.

I had a look and while it could be a little better it's not as bad as you described it to be . I notice you tend to do a lot of terraforming and quite picky about straight lines and all, but I think the route is very acceptable.

I tend to cap bus/truck connections at about 50-60 tiles. The short-range-trains i construct are much more efficient, even though construction is more expensive (especially since i'm very anal about straight lines and waste loads of cash even in the beginning).
There is also a heavy jam at the nondwood station (sp?).

Yeah, I don't take breakdowns into account (yet?), so there are definitely improvements to made in that area.

Loading the game took ages and also the acceleration button had less and less effect throughout the game. I hope this doesn't get worse in the late game. Got a Pentium 4 (2 GB RAM), which should really suffice. And then there could occur the other player who prefers 2048² with 15 AIs...

Yeah... loading isn't very optimized at this point. Below I've included a download of a version which postpones loading till after the AI has started. So you can start immediately while the AI has to load first. And the slowdown is unavoidable I'm afraid, as more vehicles are in the game OpenTTD has to do more work, and the AIs do tend to eat up quite some CPU power themselves. You can alter this by setting the construction speed of opponents lower, but this will affect the overall performance of AIs.

I still managed to get crashes for the AI, but i guess that's a doable bugfix (div by zero). I added savegames to check.

Yeah... this is really weird. Apparently AIIndustry.GetAmountOfStationsAround can return -1 for some reason...

PS: I am still experimenting with waypoints and other stuff to be able to use multiple trains on a track. That's something I didn't do before meeting the OTTD comm, so still reading and learning. But seems to be working. As said, i'm a very casual gamer, but always willing to improve. But even with obvious flaws in design i beat the AIs (not complaining about that at all, just saying!).

Yeah it seems you're doing very well . The AI still has to improve a lot and NoCAB is actually designed to be good at maps up to 512x512, for larger maps there is to much overhead at the moment although it still performs reasonably good. But once more optimizations and trains have been implemented it should be able to do better .

Thanks again for testing!

P.S. If you want, you can set the construction speed of competitors to very fast so the AI has more CPU power to work on.

[sucinum]

I notice you tend to do a lot of terraforming and quite picky about straight lines and all, but I think the route is very acceptable.
[...]Yeah, I don't take breakdowns into account (yet?), so there are definitely improvements to made in that area.

Well, i'm german
Actually i'm not too sure if some of my bridges, tunnels and terraformings ever repay thelselves (compared to a little delay on the delivery), but symmetry is also important. Eye-candy is ofc the very last thing you want to teach your AI, but i'd be grateful. If the route altogether makes fast profit, i'm fine with it and don't care going the extra mile for beauty. Performance is secondary in a game where you can get rich easily (a mediocre connection needs ~3 yrs to repay itself, a good one half the time or less). TT is not only a business game, but also a sandbox, so i think my request for nice "sand castles" by AIs isn't too far off

But the breakdowns should have higher priority, since they will make jams more severy by sticking all vehincles together. The bottleneck are still spammed stations ofc.

Thanks again for testing!

That's purely selfish, so nothing to thank for . And training starts wasn't bad either...
Otoh, my request isn't unique, so i'm happy to be able to contribute. Especially if done by mostly playing
Other AI designers are invited as well, no additional effort to include several of them in one game (if they are "casual", that is).

[Moriarty]

It's so easy to create such a list yourself that I see no need to add it to the API. Plus if you have to craete that list yourself you'll notice how big it can become, hiding such expensive functions in the api is not a good idea imo. Again, feel free to write a library that does these kind of things.

Surely if it's easy and all AI's need it, it should be in the API? Certainly from a optimisation perspective having it done in C/++ would be preferred right? As would only calculating it once, rather than having each AI do it of their own accord and in their own - probably un-optimised - way. Similarly it would reduce programmer redundency time. Also, YOU might find it easy, but people new to it may not have quite that perspective.

An no, I couldn't do it myself, I wouldn't have a clue where to start. My skills/contributions lie in other areas.

The approach you suggested would work, BUT you need to calculate or estimate the value of X well. Too low and you get inferior routes. Too high and all AI's pile up at one coal mine.

Agreed, I left that part out because that should obviously be calculated on a per-AI basis. The creation of the list however I believe should be in the API/library/whatever.

[Yexo]

Surely if it's easy and all AI's need it, it should be in the API? Certainly from a optimisation perspective having it done in C/++ would be preferred right?

Doing it in C++ would be slightly faster indeed, but since the main speed problem for AIs is pathfinding, should we also do the pathfinder in C++?

As would only calculating it once, rather than having each AI do it of their own accord and in their own - probably un-optimised - way. Similarly it would reduce programmer redundency time. Also, YOU might find it easy, but people new to it may not have quite that perspective.

These are no reasons not to do it in a library. Everyone can write an AI library and upload it to bananas. And no, I'm not going to write this one since I see no point in using it.

[Bilbo]

Doing it in C++ would be slightly faster indeed, but since the main speed problem for AIs is pathfinding, should we also do the pathfinder in C++?

Well, I thought too, that if pathfinder would be written in C++, it would be faster. But there are some disadvantages:
- In squirrel, every X opcodes the AI is "paused" and control is passed either to game code or another AI, so AI running long computations won't block other AI's or the game. It is more difficult to do this for C++ code.
- Configurability: everyone will have to use the same pathfinder and if he needs some minor modifications ... he's out of luck (or do it less quickly in squirrel and have disadvantage over other AI's). Now he can just take the one in squirrel, do the modifications and then just put the pathfinder in your AI, under slightly different name.

I think once the API and the pathfinder libraries reach some "possibly near final state", we can move some helper function to the API, so they will be performed faster and we will still retain good flexibility.

For example, _GetBridgeNumSlopes(end_a, end_b) is used in both rail and road pathfinder currently and might be good candidate for this - it is used in multiple places and it is fast (so won't block execution in C++ for too long)

[Yexo]

Doing it in C++ would be slightly faster indeed, but since the main speed problem for AIs is pathfinding, should we also do the pathfinder in C++?

Well, I thought too, that if pathfinder would be written in C++, it would be faster. But there are some disadvantages:

I wasn't seriously proposing to move the pathfinder to C++, I was trying to let Moriarty realize the disadvantages. You've pointed them out nicely.

For example, _GetBridgeNumSlopes(end_a, end_b) is used in both rail and road pathfinder currently and might be good candidate for this - it is used in multiple places and it is fast (so won't block execution in C++ for too long)

That might indeed be a good canditate, on the other hand as long as it can work in squirrel it's a very low priority thing.

[sucinum]

/edit: disregard that, i forgot to update.

I'd go for longer games if i didn't mess my city connections all the time. Trains stop at waypoints and run into each other...

[Moriarty]

Surely if it's easy and all AI's need it, it should be in the API? Certainly from a optimisation perspective having it done in C/++ would be preferred right?

Doing it in C++ would be slightly faster indeed, but since the main speed problem for AIs is pathfinding, should we also do the pathfinder in C++?

I would have guessed the pathfinder was in C/++ already, so I'm obviously not the person to ask.

[sucinum]

Here is a game i had running. I forgot to pause over night, which ruined the game for me, but allows to have a closer look at the AIs after 60 years (started 1940). It's on a 1024² map on hard diff with aviator gfx, so usual settings for me.
I played until 1955 or something, which you can guess from the amount of vehincles i had. I was struggling keeping the AIs at bay (income/efficiency-wise), but didn't interfere and could handle. But otoh, i'm looking more after straight connections then for most profitable ones.

Admiral AI has around 100 vehincles but no mentionable income, don't know what happened there. After replacing my 60 yr old vehincles, i even outperform it. Maybe that's because it crashed at some time.

The 6 NoCAB AIs are doing quite well and have around 500 vehincles each. That's not too much spamming and i couldn't see too ugly traffic jams, but some areas are really crowded (see Grontston Coal Mine or the town Dindinghead).

The AI also has issues with replacing ships:
NoCAB #5 has some longer connections using buoys and so failed to replace its ships. This is a simple error i even did myself, it happens when you order the ship to enter a depot when it doesn't find it without buoys. I didn't install a pathfinder in the game, maybe that's why. Still, this resulted in a lot of "lost" ships which the AI can't guide out of the dead end. Clearly an issue, espacially since noCAB seems to be very ship-happy. Many starts include coal ship line.

Other areas are clearly ignored, like the western corner of the map. It wouldn't be too hard to grow some of the towns there with busses/mail trucks and then build airports. I personally like improving the corners very much, in my first game (vanilla classic TT) i had each 20 concordes flying east/west and north/south corners for maximum profit (a limit of 40 applied there). This requires some improvement of the areas and i don't see the AI doing this projected. There are bus lines and growing cities, but only in those which were large enough in the beginning. Otoh, i can't really blame the AI for cherrypicking.

The game also runs very slow, but i guess i can't really complain with a map this large and 7 AIs.

Just some observations, main issue are the ships i think.

[Morloth]

Nice to see NoCAP is doing so well. It think AdmiralAI indeed did crash as it doesn't load any of its connections :/.

The AI also has issues with replacing ships:
NoCAB #5 has some longer connections using buoys and so failed to replace its ships. This is a simple error i even did myself, it happens when you order the ship to enter a depot when it doesn't find it without buoys. I didn't install a pathfinder in the game, maybe that's why. Still, this resulted in a lot of "lost" ships which the AI can't guide out of the dead end. Clearly an issue, espacially since noCAB seems to be very ship-happy. Many starts include coal ship line.

Yeah, I saw this before but didn't really understand where it came from. Your explanation seems plausible, but I can't think of a quick solution... What did you do to correct the behaviour, build a depot right next to the ships; Would that help?

Other areas are clearly ignored, like the western corner of the map. It wouldn't be too hard to grow some of the towns there with busses/mail trucks and then build airports. I personally like improving the corners very much, in my first game (vanilla classic TT) i had each 20 concordes flying east/west and north/south corners for maximum profit (a limit of 40 applied there). This requires some improvement of the areas and i don't see the AI doing this projected. There are bus lines and growing cities, but only in those which were large enough in the beginning. Otoh, i can't really blame the AI for cherrypicking.

This is to be expected. NoCAB's strategy is to start of with smaller connections (limited to 128 tiles apart) and this bound is slightly increased of the course of the game. So in the beginning, as there isn't enough money to build all of them. In effect NoCAB builds up a queue of connections to build in the future once money is available.

As income increase, the number of connections in the queue will decrease as NoCAB isn't able to calculate new routes quick enough than it can build them. This a direct consequence of more money becoming available and the average length the connections increasing (hence more time needed for pathfinding). So in the end game you will see it build less and less connections, but the connections it does build tend to be longer.

Thanks again for testing and I'm very pleased with the performance =). Hope to fix the ship issue somewhere next month once I have more time on my hands.

- Bram

[sucinum]

Yeah, I saw this before but didn't really understand where it came from. Your explanation seems plausible, but I can't think of a quick solution... What did you do to correct the behaviour, build a depot right next to the ships; Would that help?

Building a depot would only help if you do so before sending the ship to the next depot or upgrade the depot order. Otherwise the ship still has the order "go to depot <veryfaraway>" and won't find it, even though being next to one.
Also the replacement ship would have "go to harbor <veryfaraway>" as first order and would be lost again. This would also apply if the AI simply spammed a depot next to each bouy. This method would require the AI to set the next destination to a close bouy.

I personally used the bouys I had and guided the ships manually to the depot. I think this is what the AI should do as well - or simply replace ships only at their home harbors (#1 in orders).

Not too sure what's the easiest of those methods. Cash shouldn't be an issue since ships last 30 yrs (using the basic ones) and depots are quite cheap. Spamming depots wouldn't be too appreaciated for the looks, but 1-2 extra couldn't hurt.

As income increase, the number of connections in the queue will decrease as NoCAB isn't able to calculate new routes quick enough than it can build them. This a direct consequence of more money becoming available and the average length the connections increasing (hence more time needed for pathfinding). So in the end game you will see it build less and less connections, but the connections it does build tend to be longer.

Caching connections seems a clever thing to do, I just hope the AIs doublechecks before finally creating them, since i play with fluctuant economy (hard settings).

There are some unnecessary long bus connections in my current game, sometimes above 100 tiles. The busses drive through several small towns on their way, but only connect 2 larger cities. Actually a town of 700 people is enough to feed 4 busses over 50-60 tiles, so i don't really understand why the AI only goes for the largest cities. Especially on hilly landscapes, those long bus travels don't earn anything extra (not enough to cover the build costs for bridges and tunnels, that is). This also causes traffic jams, since you need a lot of busses to get a decent rating when you travel over 100 tiles away.

Another odd thing are the signs the AI uses. I have seen a lot of signs reading "T" or "!!!" all around since i play with noAI, mainly next to stations. What are they good for? I have seen an AI prepare a place for an airport and placing a sign there. I simply bought a share of land to prevent this (important city for me where i can't afford sharing/losing passengers), so just wondering. I wouldn't mind if the AIs used signs with real text for debugging or something, but those are simply puzzling me...

[Morloth]

Yeah, I saw this before but didn't really understand where it came from. Your explanation seems plausible, but I can't think of a quick solution... What did you do to correct the behaviour, build a depot right next to the ships; Would that help?

Building a depot would only help if you do so before sending the ship to the next depot or upgrade the depot order. Otherwise the ship still has the order "go to depot <veryfaraway>" and won't find it, even though being next to one.
Also the replacement ship would have "go to harbor <veryfaraway>" as first order and would be lost again. This would also apply if the AI simply spammed a depot next to each bouy. This method would require the AI to set the next destination to a close bouy.

I personally used the bouys I had and guided the ships manually to the depot. I think this is what the AI should do as well - or simply replace ships only at their home harbors (#1 in orders).

Not too sure what's the easiest of those methods. Cash shouldn't be an issue since ships last 30 yrs (using the basic ones) and depots are quite cheap. Spamming depots wouldn't be too appreaciated for the looks, but 1-2 extra couldn't hurt.

At the moment I use the automatic replacement options, which clearly makes a mess in the case of ships. So I guess I will have to switch to manually dealing with this. Thanks for the tip .

As income increase, the number of connections in the queue will decrease as NoCAB isn't able to calculate new routes quick enough than it can build them. This a direct consequence of more money becoming available and the average length the connections increasing (hence more time needed for pathfinding). So in the end game you will see it build less and less connections, but the connections it does build tend to be longer.

Caching connections seems a clever thing to do, I just hope the AIs doublechecks before finally creating them, since i play with fluctuant economy (hard settings).

It does actually, it caches them in order to make decisions what the build given the current amount of money. In other words it helps the planning process. But right before each connection proposal is presented to the decision making process the costs are recalculated so no worries .

There are some unnecessary long bus connections in my current game, sometimes above 100 tiles. The busses drive through several small towns on their way, but only connect 2 larger cities. Actually a town of 700 people is enough to feed 4 busses over 50-60 tiles, so i don't really understand why the AI only goes for the largest cities. Especially on hilly landscapes, those long bus travels don't earn anything extra (not enough to cover the build costs for bridges and tunnels, that is). This also causes traffic jams, since you need a lot of busses to get a decent rating when you travel over 100 tiles away.

That's actually the point. If more busses can use the same road it means you will spread the fixed costs (i.e. infrastructure) over more vehicles and get a higher per-vehicle return in the long run as the fixed costs are spread out over more vehicles. In other words it's more efficient . I let the AI calculate the expected incomes based on travel times, fixed and variable costs and based on the replacement time of each vehicle (this is used as the horizon) the most netto profitable connection is build. It takes things like slopes, etc. into account when checking how long a vehicle takes to cover a given distance. So I have faith that NoCAB only chooses the most profitable connections based on above criteria from its queue. I have noticed however that some AIs like to use drive-through stations in cities which can create massive congestion. I'll address this issue in the next version and there are some improvements in the congestion detection algorithm in NoCAB to handle this issue. But all in good time .

Another odd thing are the signs the AI uses. I have seen a lot of signs reading "T" or "!!!" all around since i play with noAI, mainly next to stations. What are they good for? I have seen an AI prepare a place for an airport and placing a sign there. I simply bought a share of land to prevent this (important city for me where i can't afford sharing/losing passengers), so just wondering. I wouldn't mind if the AIs used signs with real text for debugging or something, but those are simply puzzling me...

Yeah... those should be removed from published versions. In the case of NoCAB it's simply debug information, it signals that terraforming has failed in that area (probably in anticipation of creating an airport there) or that the airport couldn't be build after terraforming (e.g. not enough money, local authority getting mad at terraforming, etc.).

Thanks again .
- Bram

[sucinum]

That's actually the point. If more busses can use the same road it means you will spread the fixed costs (i.e. infrastructure) over more vehicles and get a higher per-vehicle return in the long run as the fixed costs are spread out over more vehicles. In other words it's more efficient . I let the AI calculate the expected incomes based on travel times, fixed and variable costs and based on the replacement time of each vehicle (this is used as the horizon) the most netto profitable connection is build. It takes things like slopes, etc. into account when checking how long a vehicle takes to cover a given distance. So I have faith that NoCAB only chooses the most profitable connections based on above criteria from its queue. I have noticed however that some AIs like to use drive-through stations in cities which can create massive congestion. I'll address this issue in the next version and there are some improvements in the congestion detection algorithm in NoCAB to handle this issue. But all in good time .

Actually i'm not too convinced of that. This only works if there are no jams, and especially on bridges, a single breakdown can make a huge loss if 10+ busses are blocked. Also there tend to be large chains before stations, because sooner or later, most vehincles will be stuck together (especially with bridges in the connection). This is a loss of time i won't have with 2-3 busses per station. This really depends.

Just for good measure, i have uploaded yet another savegame. This time 512² and other settings as usual. You can see a lot of different bus connections there.

I have several small ones and most of my busses net 7-9k a year. That's not too spectacular, but actually that's the best performance. (Just found out that i will never get a score of 1000 without inflation on unless i sell all busses and trucks, because i will never earn 20k a year with them. How lame.)

There is one real good bus connection the AIs have set up: AI#6 between Drunfingley and Nundhead Cross. I have an own connection close over a shorter distance and actually have worse ratings and less income/year per bus (Built it because of a subsidy and the spot wasn't too bad actually).

Most other busses of the AIs earn less than 5k a year and are jammed very often.

For example Finnbridge clearly suffers of jams, even though there would fit another bus station (I guess the AI doesn't use building with Ctrl).
Even worse in Sondhattan, and the AI could easily replace the depot there with a 2nd bus station. Sondhattan is connected with Slunnville, which also suffers badly from traffic jams. All busses in this connection make 3-4k a year, which is clearly too low. A 2nd station in both towns could improve the situation a lot, but as it looks, the AI doesn't remove town buildings or stuff it has built previously (like that depot). Or uses "green belts" for the local authorities.

So in theory you are right and there is a nicely executed connection in my game. But the majority is jammed and so has low performance. To make profit out of a distant bus connection, you are bound to build 20-30 busses (because of shared building costs as you explained and also to get decent station ratings). And that can be hard to manage.

I think it's very hard for an AI to "think" of all bottlenecks which can cause a bus connection to jam. Especially since any other AI (or I as human) can share the roads. So i'd suggest to stick with shorter distances.