Trains: semi-automatic mode and autopilot

[XKnight]

Finally with 0.13.17 we can slightly control our trains using circuit network, but I think we need more...
In my opinion, devs made a great job by constantly improving train pathfinding algo, fixing bugs in it, and adjusting algo to our needs, but it is extremely hard to satisfy all users with a single algo.
That's why we need to control it.

The only reliable way to affect pathfinding algo is by changing penalty of some blocks (yeah, and it is not possible at this moment), but what does this penalty mean? For the regular player without any knowledge about internal constants inside PF algo and its implementation details these constants just a dumb numbers.

That's why I will not propose to add any penalty items, but I would like to propose to move current PF into a separate mode, like "smart autopilot". Also, new "dumb autopilot" should be added for all other needs. In this new mode train will select any green path that is available, so players should manage all signals using circuit network.

Obviously, it should be possible to switch different autopilots using circuit network, for example when train is passing by a rail signal.

For example:
if you want to create a train depot/waiting bay you should switch train into a "dumb" mode when it enters depot/bay, then manually (this means using rail signals) conduct this train to a free waiting stop, and switch it back into a "smart" mode when it leaves the depot/bay.

This mode will allow players to tweak PF algo without its direct modification, also we will receive a new toy .

Any ideas?

[ssilk]

I think this is to complicated and you will not see, if you programmed it wrong.

My counter suggestion: I want a Really Smart Pathfinder, cause a really smart pathfinder is a pathfinder which knows, where to be smart and where to be dumb.

[quadrox]

I don't understand your suggestion - I am usually not stupid, but it seems to me you skipped a few steps in your explanation:

The only reliable way to affect pathfinding algo is by changing penalty of some blocks (yeah, and it is not possible at this moment), but what does this penalty mean? For the regular player without any knowledge about internal constants inside PF algo and its implementation details these constants just a dumb numbers.

This mostly makes sense to me.

That's why I will not propose to add any penalty items, but I would like to propose to move current PF into a separate mode, like "smart autopilot". Also, new "dumb autopilot" should be added for all other needs. In this new mode train will select any green path that is available, so players should manage all signals using circuit network.

How is this different from the current autopilot? How does it relate to the penalty system you proposed above?

Obviously, it should be possible to switch different autopilots using circuit network, for example when train is passing by a rail signal.

Ok, that sounds good, but I still don't understand what the proposed new autopilot does different.

For example:
if you want to create a train depot/waiting bay you should switch train into a "dumb" mode when it enters depot/bay, then manually (this means using rail signals) conduct this train to a free waiting stop, and switch it back into a "smart" mode when it leaves the depot/bay.

I really don't understand this part. Why would you need to switch modes when entering a depot? what would the different modes do different? Please explain in more detail.

[XKnight]

More explanations to the God of Explanations.

Autopilot usually has its own criteria which should be optimized, in our case this metric is a travel time (and maybe traveled distance).
Although each autopilot wants to optimize its own metric the whole system may not reach its maximum.

First example

Two trains on the same track: first train has 5/5 wagons with iron ore, and second train has 0.5/1 wagon with blue circuits.
First train has low speed because it is heavy, and second train also has low speed because it cannot outstrip the first train (second train is faster but it is not super fast, and overtaking path also is not very large).

What we have? The system can not reach its global maximum, because first train "think" its metric is the most important for everyone.
But honestly, who cares of several seconds delay for ore train when blue circuits train is delayed by tens seconds?

This problem can be solved in two ways (building second rail or making overtaking path longer is out of scope):
- make current PF algo genius
- or temporay disable ore train's brain during overtaking path, and force it to sacrifice its metric for the global purpose

Second example

Sometimes it is needed to make something irrational (at the first blush) to obtain better results.
Example: crossroads.
Very-very smart autopilot will never wait on the crossroads. Why? because it is always possible to reduce train speed before crossroad in order to pass this crossroad with non zero speed. As a result - faster accelerating.

How this can be solved?
- improve current PF algo, and sell Factorio with a small elf forecaster inside (joke).
- leave this problem to players

Last example

Waiting bay. Lets take for example a distributed factory and a smelting area. It has one drop off point (for different ores) and several pick up points (iron plates/copper plates/stone bricks). Also you have 5 supply trains assigned for iron ore, 5 - for copper ore, and 5 for stone.
How waiting bay will look in this case?
- simple line; first in - first out. This is not very good solution because such system can't adequately react on signals from the outter world (increased/decreased demand of some resource). Waiting bay length: 1x15 trains.
- single train - single waiting stop. Quite nice, but if you have 5 trains per each resource (iron ore/copper ore/stone) you will need at least 11 different waiting stop, only in this case you will be able to switch to different resource. Waiting bay length: 15x1 trains.
- single resource - single waiting line. In this case you will need to manually control destination line for each train (this is possible to do using pass-thru station). Disadvantage of this solution is longer waiting line. Waiting bay length: 3x5 trains.
- single resource - single waiting line with limited length, all trains above specified threshold will go into the shared circular waiting bay. Waiting bay length: 3x1 train + 12 shared storage.



Obviously, there is no PF algo that will be suitable for this system, because shared circular bay breaks all train's personal metrics... and makes resulting system more valuable.

Yes, I known that this problem can be solved in many other ways but it is used just for example.
Possible solutions:
- using separated smelters (partially)
- using separated drop off points (partially)
- preventing train spamming
- ...

Obviously, it should be possible to switch different autopilots using circuit network, for example when train is passing by a rail signal.

Ok, that sounds good, but I still don't understand what the proposed new autopilot does different.

The only difference - is hiding PF's implementation details (penalty system) from player and adding possibility to affect train's brain, that's all.

P.S. sorry for bad pictures

[ssilk]

Nice explanation, much afford.

Your first example shows a problem which can be solved with the latest patch about train signals (better: I think/hope so).
Instead of making the trains smart make the train-control smart: Switching the slow, long train to the side-track if a fast train follows seems to be possible now.
The same is true for the last example.

And for the second example: Even in real world such problems cannot be solved by teams of genious people.
And it is not needed, cause in realworld we have bridges/tunnels.
That will be much easier and more understandable, than any of the discussed changes.

[XKnight]

Your first example shows a problem which can be solved with the latest patch about train signals (better: I think/hope so).

Yes, it will. Heavy train will select free path, but it will not wait there. So you will have to
- open waiting path, close main path
- heavy train is going into waiting path
- open main path, close waiting path
- wait until fast train pass through this block
- open waiting path

I would say: this is a train trap

The same is true for the last example.

Unfortunately, in this case the answer is no. Train will never select cirular waiting bay, because it has higher penalty than just waiting on the main track. The same answer is for trains inside this bay: these train won't circulate because every train wants to exit the bay as soon as possible.

And for the second example: Even in real world such problems cannot be solved by teams of genious people.

That was just an example, also, real world has a real problems and most of them are thrown away in Factorio.

[person3triple0]

I think this could also be a balancing issue with a heading of Train Prioritization.

There could be a way to prioritize which trains leave first (maybe using the coloring thing? or some other method).
Reasons it would be useful:
• Base is under attack, but you're on the opposite end
• Set your train to highest priority (emergency), all the other trains put an effort in to either get out of your way, or let you go first, even if it means going in a direction they normally wouldn't go.
Alternatively:
• Setting high level, high cost objects to a low priority, plates to middle, ores to high, coal highest(not emergency, though), so that the factory continues to run if a low level train runs out of fuel and blocks a line.