Train AI is not that smart!!!

[zingo2]

Hello All!

I made this topic on Steam forum. [not allowed to link on first post]



Hello all!

There is a new problem. I have a 4 section roundabout.





Now as you may all know 2 section roundabout is almost flawless. 1 rail in 1 out- it works. The problem though with 2 way rails they can only sustain up to 40 trains +/-. More than that and you will have major train jam.

As there is no such thing as 3 way rails(in loop design) next common thing to do is to upgrade to 4-way rails. 2 in 2 out. Thus giving you more throughput.

BUT!!! Here is 1 major BUT!!!!!

Train AI or logarithm only prefers the most fastest route- not the smartest but the most fastest- shortest route. And Houston- we have a BIG problem!!!!

As the train Logarithm prefers 1 setting- it does not valuate the smartest option. it will wait in "Que" manner until its turn to go and you cant do nothing about it.

As you see this is almost a perfect 4 way roundabout- with 4 circles. Giving the limitation of the AI for trains at most you can only see 3 trains in roundabout. As they ALL prefer the inner circle. Even before the roundabout if its clear- it will prefer the inner circle. In real life smart logarithms would allow multiple trains in roundabout but here- no. As factorio is in EA there is hope for smarter train AI.

As that currently the most circles you can have are 3- not 6 nor 1 but 3. Antyhing more than 3+ will result back to 3. Say 6 track roundabout will act as a 3 circle roundabout. And you cant do nothing about it.

[Shokubai]

Can you upload your save game so I can see your logic?

Trains are dumb but there are ways to use their stupidity to your advantage.

My initial thought is that you have given your trains too many choices. Dumb it down...use less crossovers.

[theRustyKnife]

Not trying to be mean or anything, but I think you meant algorithms rather than logarithms. I highly doubt the train AI actually uses any logarithms...

As for the original topic: yes, trains are sometimes very dumb and I wish some more work would go into their AI. I'd also like them to be more predictable. In the current state, stuff like waiting bays is almost unusable because of the weird behavior...

Cheers TRK

[zingo2]

Haha My bad!

Algorithm it is.

As this is not my second post here I will link the discussion topic from Steam to here as many of the answers are answered on Steam page. Not a fan of dual linking discussions but for now it is what it is.

GO TO STEAM:
http://steamcommunity.com/app/427520/di ... 092267028/

I would appreciate if some of the Developers could comment on this topic- say- put things what they are now and what can we expect in future. Maybe a Developer has a better 4way or 6 way roundabout Blueprint he wishes to share with the rest of the community that eliminates the problems I have come upon.

[Xeteth]

This is why roundabouts are bad practise. I will get flamed for saying this; which is fine.

My suggestion is to use T-Junctions instead; you can still use a loop based system but just stay away from roundabouts. My second suggestion is to remove the lane switchers - they are often not needed and are quite taxing on the train pathfinding.

Edit: Here's an example of a 4-lane T-Junction.

Edit 2: I know it's missing a chain signal in the bottom left, make sure to fix that if you use it

[Shokubai]

A lot of info missing here to answer your questions.

Length of trains
Distance of travel
Main line path through your junction (predominant direction)
Signaling complexity
Path complexity outside this junction

You have asked for specific help but are only giving general info. Upload a save so we can see more detail about the issue and we may be able to solve your problem or at least provide some workarounds.

As it is...The rail in your image is over-complicated for those "dumb trains".

Also you keep touting this as "ideal" but it is, in fact, at the heart of your issue. You are allowing trains to spend too much time in your junction by giving them too many choices(very long ones at that). Remove the outer ring and add strait track across the middle with chain signals.


Let me approach a solution for you that may help you. Ideas to consider (not necessarily together).

a)Consider moving away from 2 way main lines. A one way loop design may help increase your throughput.

b)Consider express-laning. This idea simply meens DONT GIVE ALL TRACKS EXITS at ALL JUNCTIONS. Allow trains at the furtherst points to access a rail that gives them priority and skips traffic...A bypass if you will.
This was from another thread I posted in where i talked about a similar idea. Essentially, a train going only one junction away would ALWAYS use the outer lane. A train going 2 or more junctions away would ALWAYS use the inner lane. This obviously doesn't fit all problems, especially if all trains go to one location but it emphasizes my point about fewer choices for dumb trains.



c) Consider segregating your network. Why do all trains need to share all rail anyway? Sometimes it's best to put some things on their own network. Then you can cross a track if needed without crossing at or near a junction.

d) Optimize your trains. make sure you arent sending trains that are longer than they really need to be. Make sure trains aren't traveling more frequently than they need to be. Plan your stations ahead to avoid congestion.

e) Early exits and long mergers. Where space exists, using extra track to create merge lanes and exit lanes before and after a junction can help reduce stress if of sufficient size.

d) Recognize your bottleneck may not be the junction alone. Stations close to a junction, wrong priority on merges, or 100 other things could be triggering a caterpillar effect through the junction. Fix issues away from the junction and re-evaluate the junction for changes.
[img]do

[zingo2]

Hmm Okey For the fun of it

DropBox link:
https://www.dropbox.com/s/z1qsxvreplbt6 ... 3.zip?dl=0

Its a dedicated server with the mods listed here:
http://steamcommunity.com/app/427520/di ... 117362603/

Though I doubt a save game will help but then again I could be wrong.

As I stated before original discussion is on Steam here:
http://steamcommunity.com/app/427520/di ... 092267028/

Anyhow:

Length of trains: 2-4-0
Distance of travel: varies(short-medium-long)
Main line path through your junction (predominant direction) (2 way in 2 way out /// 4 way tracks // loop design)
Signaling complexity : Max complexity- Signals are designed for 2-4-0 trains.
Path complexity outside this junction: Complex.

+ The map is not finished. We did a major overhaul. When we got to 40 trains we started having train jams that's why we took the approach to upgrade from 2 way to 4 way tracks. The 2 way roundabout didn't cut it anymore.
Possible trains when completed- aprox 70 to 80 when finished. + extra 10 to 20 supply trains form various outposts.

I have a 3 way T-junction here- though not pretty it works 100%


Do you think a 4 way junction would be best answer for now? Like replacing roundabouts in favor of 4 way Junctions?

I noticed that too- that "in its current state train AI is very dumb for the given design" As you can easily remove 1 or 2 inner rings from roundabout and it still works.

Soo yea- Im open for suggestions!

[Shokubai]

Edited post above while you were posting.

[zingo2]

Haha Now thats what im talking- people who actually know what they talking and fast!

Okey lets see:

This is our map



There is A huge train station with Depo(Depos are standart for all factory complexes) in our MAIN BASE.

Then there are: Stone/ Copper/ Iron/ Coal/Oil/ Modules/ Engines/ Green/ Red/ Blue and rocket Bases (or works or factory complexes(at least 100 factories)) The 3 chips together will be separated in future- though they work they are a bit tad too close together.

4 lanes were implemented for the central area- where there is heavy congestion of trains. We can safely build 2 way rails outside the central without any problems.

Okey to Answer Shokubai :

a) I agree but listed above and 2 way lines were not cutting it anymore so we upgraded to 4 way- the train jam went away immediately.
b) This is a great idea however building such a thing would require lots of real estate. While i think it may be possible The design would be very complex to do simple tasks. Keep in mind there are other people on the server! They have difficulties grasping the 4 lanes now imagine- express ways- cut outs etc etc
c) Because the main base is in central- or parts of it. Basically Whole base was split up in these factory complexes that are connected with lifelines- via trains. As some factory complexes share a connection with multiple other complexes/ cells- its all connected/ interlocked- you name it- 1 living Gigantic Organism. You just have to feed it 4 raw resources and magic happens.
d) Optimizing trains/ timings is a constant battle on daily basis. So check.
e) Already done (i think).
d) I agree. So far everything is working like a dream. ONLY except for AI dumb behaviour of always choosing the shortest route!!! Not the smartest or effective but shortest! So far the Biggest Bottleneck in this connected system is roundabouts themselves.

[mooklepticon]

This is why roundabouts are bad practise. I will get flamed for saying this; which is fine.

My suggestion is to use T-Junctions instead; you can still use a loop based system but just stay away from roundabouts. My second suggestion is to remove the lane switchers - they are often not needed and are quite taxing on the train pathfinding.

I'm only on 2 lanes, but I've gone to this. T Junctions and Cross Junctions only. All loops are done on the side, not in the main track. For example, I'll do a cross junction and one of the exits of the cross is a loop, and only a loop. It's extra track, but it prevents a lot of headache.

Haven't needed 4 tracks, yet, so I don't know if it scales.

[Shokubai]

Currently playing your save. ...Kinda starting over on your rail. Gimmie a bit.
on my A) item about two way I didn't mean two lanes...I meant TWO WAY with any number of lanes. The idea being that trains can only go one direction period. They must then make a big loop to get where they are going but all trains are always going the same direction. In theory this should handle more traffic overall causing less jams but creating greater travel time.

[Shokubai]

I didn't spend a ton of time on this but I have some tips for you. I was half a mind to just redo everything but then I know that I learn best by doing so here goes.

First, I wanted to play with a one way loop and your base is set up nicely to facilitate these. Have a look around your iron works for an example. I used two lanes and allowed all trains access to both. You could refine this by fiddling with bypasses and such or adding extra lanes. Overall I was happy with it but there is plenty more you could tweak here. Don't expect it to be pretty, I wasn't going for symmetry and I didn't spend time optimizing.

Next there are some really BAD spots that need a rework. One of the worst is this area. Frankly I think you should merge the In and Out paths for these two bases into one. This will lessen the impact on your main line. Also I am not sold on needed 4 lanes through here or giving access to all 4. You look like you are probably 75% to your goal quantity of trains and I just don't think its necessary. One Idea I considered was bringing the In/Out to the loop on the right and shifting the junction South a bit where the Southern area merges in but you would need to evaluate the paths of all the trains a bit.

JAM.png (51.12 KiB) Viewed 13246 times

This is kind of insane and keeps trains in the junction WAY too long.

slightlyinsane.png (12.27 KiB) Viewed 13246 times

when this would do the job. If I expanded this I would add Right turns earlier in the junction to bypass the heart of things.

This.png (9.42 KiB) Viewed 13246 times

Place signals very close together in Acceleration/Deceleration zones such as exiting/entering a junction or station. This helps more than you think. You generally need more signal density throughout. Maybe half a train length between signals.

Also, remove all the lane shifts. These really don't help and just make new places to force a train to stop.

Lastly, something else to think about. Nothing says all trains have to use all lanes. You may consider creating specific lines for specific chains. A train from the iron ore mine will not share a track with a train from copper ore. They may have parallel lanes but are not given the option to get into eachother's lane. This may work best in a 1 way loop scenario unless you can keep your number of lanes down as two way could take up a lot of space.

multiline.png (7.9 KiB) Viewed 13241 times

[zingo2]

Hey Thank you Shokubai for the work you did!

So I fired up your work and... hmm .... yeah....

What you basically did was dumb it down to train AI level However there are some problems.

*) The T-junctions you created are basically the same 2 way roundabouts when we had 2 way lines. The average train time is the same- though it seems faster the train spends the extra time elsewhere. Though that 1 circle pass is interesting but it has a limit of how many trains it can sustain. If we add 50 more trains here- it will get over populated. Though it clearly demonstrated how inefficient(dumb) the train AI is.

*) 4 lanes was designed with the idea of supporting at least 120-150 trains as with 2 way we hit a ceiling of 75. Maybe that's the only reason we went with 4 lanes.

*) I agree- the 4 way roundabout- while a Modder could write a code for smarter AI- at current state Factorio Trains are just TOO DUMB to utilize it - that's a shame really...

*) Yeah placing signals at half the train length increases responsiveness- like similar with PC architecture- via pipeline. The pipeline has a limit but there are tricks (HT and stuff) that can tweak it.

*) On paper Lane shifts work- in factorio- not yet. Also a pitty. Though you need them not so close to intersections.

I talked to the guys of the findings on Steam and here.

** We like the idea of Express lines- Its possible to make some of them connecting specific "cells" thus pulling some amount of trains off the main grid.
** We also like the idea of replacing roundabouts with 4 way T junctions(cross section)
** There are some sections we could improve based on your findings(like with the iron works in/out)
** There is also an option to move raw resource Bases at the edges and make a round express way leaving the center for the rest- thus dividing the network in 2 however future problems may rise of sustaining larger amount of trains. Still debatable.

Overall thanks for the ideas Shokubai Some of them are very good and could be easily implemented.

I wish a Developer would answer whats the roadmap for future train AI progression- so to make right calls (in which direction to build as human mind is full of wonder). Or maybe a Modder can write a better code. cause it does not really matter if a train looses 5 or 4 seconds here or there but if it makes a smart decision- that's a game-changer.

Anyhow If any of you guys come up with a working solution for 100 trains- post here.

Once we hit 100+ trains Ill update this thread how things are progressing and how the train traffic is coping.

[Shokubai]

You are welcome and sorry for the incomplete example on your save but it was just more work than I wanted to do to keep going. Still I am glad it gave you some ideas on where you might do things differently. At least I think you have a way ahead.

There has been a lot of discussion recently about improving train behavior such as adding a path signal to the game which would allow multiple trains inside a signal block as long as they both can find a path through the block. You should check out the Ideas and Suggestions threads and weigh in or offer your own ideas. This is an area where I feel Factorio needs a lot of work.

You may want to send ssilk a PM. He is a forum mod here and would probably be best able to tell you what "the plan" is though I think train logic is generally low priority atm. Here is one of his train networks from the Pathing thread. (Notice the lack of 4 way junctions and use of T or 3 way junctions)


p.s. Just as a note on terminology. You keep saying 4 way when you mean 4 lane. A 4 way T junction would be a + junction (N S E W = 4 Way). A T junction is by definition 3 ways though it may be any number of lanes.

[ssilk]

Hehe. I need to add here, that I use some circles, cause they are useful, when you need to return with the raillayer. And in the long run, they made me the most problems, because of the continuous deadlocks they provoke.

But in general I ask myself, why other players have always 4 lanes. I nearly never needed it. Especially with long tracks and - even when there where traffic on it with more than 10 trains per minute! But I need it at the ends. Before and after the switches.

That is an interesting solution and I will explain it a bit better here.

TL;DR: The faster your trains can go in average, the less tracks you need. Add more tracks before or after a switch.


One of the best examples, that everything can run over one track per direction is in the center of the above picture, a little bit to the left. Here is a zoomed in pic of that part:

Screen Shot 2016-06-21 at 14.43.57.png (7.3 MiB) Viewed 13195 times

(you can double click to enlage the pic)

This is part of the Zentral Depot Router (Zepot). Here the sorted items are loaded to trains for all the sub-factories outside. Something I would never build again, but it was quite interesting to see, how that structure could handle the rising traffic. The interesting part are the rails in the red box. Trains come in from right on 3 tracks. And then I made a fuckup: I left not enough space. Only one track goes to the left where it splits again into 3 tracks.

One would now assume: Ha! The bottleneck is this part with only one track.

It's not! Indeed (after some tweaking) there is no bottleneck at all, even with 20 or 30 trains per minute. When I need to compare, I need to say: The bottleneck is more at the beginning or the end of the red box, but the trains in the middle jams only cause of this bug, that the trains sometimes stop, even if the rail before is free.

Now, this remembered me to something. This is the same, as the pressure in a arosol tin:

(From https://de.wikipedia.org/wiki/Str%C3%B6 ... nd_Venturi)
The translation of this is: At the closest part, the pressure is the lowest, cause the speed is the fastest.

Now: When you rotate this picture by 180 degrees you see the similarity with the above part of the factory!

That is of course not always true (there are always exceptions), but it is enough valid for me to make it a rule.

[taiiat]

this is basically answered already, but i just want to use some roadwork theory here.

adding more lanes to your road, does not make cars go faster. Civil Engineers (or people in charge of Cities that actually know nothing about roads) tend to make this mistake. 'slap more lanes, that will fix it!'. more lanes usually either makes no difference, or actually makes the problem worse.

because, the only thing more lanes does, is increase the buffer capacity of every N length of roadway.
when planning out large road networks, the logistic problems you're trying to solve are ceases of movement, not vehicles being on the roads. you want to keep traffic constantly flowing, if you do that, everything is at maximum efficiency (or close enough to).
in heavy traffic situations, you're never going to have all vehicles at their maximum speed. but that's okay. things get there faster if they consistently move at half their speed, rather than flip between maximum speed and not moving many times along the trip.
the biggest time increase for driving from A to B is not moving at all!


so then, some quick Bullet Points of important things to have in Roadways to maximize traffic flow.

• Turning Lanes are critical. vehicles that aren't planning on turning, shouldn't be forced to wait for just the one that is. turning lanes gets the vehicle off the high speed travel area before it comes to a slow/stop in order to make the turn.
• don't give too many options. more ways to cut between roads actually hurts road networks. they cause varying speed chaos.
  • do give a few though! don't make a vehicle travel halfway across the galaxy just to like, turn around. but keep these options cordoned to only vehicles that are in the general area - have some options nearby, but don't have so many that you just end up with an octopus monster.
• A/B/C Roads. this means plan things out in separate road 'Tiers'. 'A' Roads are the mega highway that is constant high speed travel over long distances. 'B' Roads are intermediary, they route traffic that's going to a relatively near destination - generally traffic that's going to somewhere that's outside of a stones' throw but not quite many towns away. 'C' Roads are basically the smallest Roads you can have, for connecting local traffic to the 'B' Roads.
  • splitting up traffic when possible is also applicable here.
• smooth turns result in higher traffic flow. sharp turns cause traffic jams.
• more complicated roads are not necessarily better roads.
• one way loop systems are a lot more effective than the seem at first glance.

now, a couple of these don't really apply to Factorio, due to Trains having some space magic to them. sharpness of turns don't really matter, Trains can take turns at top speed without worrying about Derailing or even having to slow down.
and for those sorts of reasons, Turning Lanes are also not that important (but in a truly high traffic scenario, could become useful - may be worth considering since you're planning on having hundreds of trains).

but for the most part, treat your Railway networks just like you'd treat a Roadway network. you know that your Citizens will aim for the route that is either shortest distance, or shortest travel time given maximum speed at all times.
you do indeed partially end up planning for stupidity. but you know it's going to happen, so you plan for it.


yes... i did spend quite some time helping people with Roadwork/Traffic Theory on r/CitiesSkylines.

[zingo2]

Alright! After Much of frustration- some cold beers and a shower- almost smashed monitor(and surroundings) and even a lot more of beer(Charley is that you???) i have come to a "partial" solution!

(Its a joke of course but this did cause major headaches and a lot of time)


So lets get it on!

Once you reach a point in Factorio where single rails dont cut it Its time to upgrade to 2 way then 3/4 way.

I. 2 Way Roundabout // this is 99% good roundabout up to 40 trains +/- (your mileage will vary on the train design and signals you have)




@@ +++ Its a good intersection for 2-way rails- very module- easy to build and very responsive
--- It looses effectiveness the more trains you have.

II. 4 Way Roundabout // This is a very complex design- it works on paper but in Factorio- this is a pass... Train AI cant handle it (0.12 version)




III. So I went with this design. In reality however i do not recommend this. While it looks pretty it is the slowest roundabout of all my tests.




IV. Then I removed the rings altogether and this is what i got. However you need a connection somewhere on your rail for UP/DOWN rails to connect. Its not even a roundabout but it does lets you switch lanes very fast- very efficiently. I just couldn't work out an efficient way for UP/DOWN tracks to connect in a productive manner(And i tried X, C and half loop sections- all reverted for trains getting stuck or loosing momentum)




V. I tried many designs posted in factorio . Blueprints website. One thing i realized- not all designs are functional. I mean on paper- in picture- they look logical and pretty but I tried like 10 designs ending in failure. Though this one actually is functional. Credit goes to OnlyCeraphix I took his design- modified it a bit- signals here and there- extra tracks and curves here and there- and this is the first "proper" 4-way roundabout/intersection that works!




@@ +++ Compared to all designs I tried this one so far has the highest success rate for 4-way tracks. The AI at some points can fit 4 trains simultaneously here.
--- Haven't tested the full limitations of this design but it should handle up to 80 to 100 trains(test results pending)

VI. This was addition to IV. This is the best i got However it will clog at one point and it will fail. I tried placing them very far- then very close- even tried merging them- no avail. If you can come up with a connection design that goes together with IV design- I would love to see that.





----------------------

So yea What all of this means:

* Its better to have 1 intersection with turn signals then have 2 separate- an intersection and turn/connection loop. Ive tried many- didn't come up with a successful design. Best i got was VI. You can do some magic with signals but it will lower the train speed.
** The roundabouts do work but you need to design your factory in a logical manner. Better Utilize roundabout quarters then UP or DOWN or a loop. For example place your Iron factories DOWN and to your right place your other factories that need that iron. You will Utilize a Quarter of the roundabout with flying colors.
*** If you manage to sustain your many trains at high speed (above 200) you need less trains and the whole train traffic is smooth. Design your railway so the TRAIN AI can have advantage.
**** Train AI is very dumb- don't expect miracles. If something is not working dumb it down to TRAIN AI level.
***** Pay very close attention to your signals. Signals can boost or decrease your train speeds- placing them correctly is very important!
****** Try to avoid using too many roundabouts. While there are cases they are needed and there is no other option any T-junction works- AND it does not matter if its 1/2/4/6/8 way junction. It will work regardless.

[Qon]

Anyhow If any of you guys come up with a working solution for 100 trains- post here.

If the goal is throughput and not a certain number of trains you could make them longer instead of adding trains. More cargo wagons means more capacity per train. More locomotives means higher top speed and acceleration.

[zingo2]

Anyhow If any of you guys come up with a working solution for 100 trains- post here.

If the goal is throughput and not a certain number of trains you could make them longer instead of adding trains. More cargo wagons means more capacity per train. More locomotives means higher top speed and acceleration.

In this particular case the Trains are 2-4-0. If we went for longer trains we would have to redesign all signals and all roundabouts/ T junctions/ stations/ Depos- basically everything.

You choose 1 train design and design railway network around it- that's how that goes!

The goal is both! Keeping smooth throughput by adding more trains. I mean this case is 2-4-0 But maybe one of you guys has 1-3-1 maybe 2-8-2 etc etc. And your setup works very well with 100 trains. So don't be shy to share your concepts As any concept can be adaptable to any design(mostly)!

[Qon]

In this particular case the Trains are 2-4-0. If we went for longer trains we would have to redesign all signals and all roundabouts/ T junctions/ stations/ Depos- basically everything.

You choose 1 train design and design railway network around it- that's how that goes!

Why would you redesign the rail system? Having blocks shorter than the train doesn't really hurt. If your junctions are correctly made then train length is almost completely irrelevant. With less trains your junctions should work better. The only thing I can understand why you would redesign is stations. But a train with twice the amount of wagons can unload in your current stations if you add another train stop so it unloads half the train at a time (not optimal but might be good enough in some systems).