Things about the AI (artificial intelligence) component that are not prototypical.

JonMyrlennBailey

Well-known member
As most know by now, AI doesn't mimic human train control well or even the computerized autonomous railroad operations as on some modern inner city rail transit systems.

1. braking/acceleration is often too sudden given the size/weight/speed of the mimicked railroad vehicular equipment that Trainz is generating images for
2. the unnatural reduction in throttle/train speed when passing a track mark (waypoint) while executing "Drive Via": an autonomous car would not slow down while passing a waypoint while it tries to calculate the next command in its drive schedule, neither would a jetliner in autopilot mode
3. the fact that this system sees RR signals too far in advance, ideally the AI component should be reengineered to only see an upcoming signal within the range of vision of a human train operator, perhaps 500 to 1,000 feet and react to it in the manner a human train controller would react to it; a one one-mile-long freight train should not slow from 50 mph to 25 mph in 5 seconds or less when it's human at the control stand of the lead engine observes a yellow condition: a human controller will have perception time, reaction time and the train, following physics laws, will require a certain braking distance given its mass, grade, track conditions and speed
4. diesel locomotives under AI control unnaturally rev high continuously once cruising speed is reached, a human operator, on the other hand, will adjust throttle position to keep the train at set speed and you will continually hear fluxuations in engine pitch: when the real-world diesel engine hits a target speed of 50 mph on a level section, the engineer may close the throttle all the way and only increase it as the train slows down to bring it back to target or control speed again: having been a passenger on d/e trains and seated close to the engine, one can hear the variation of engine pitch as the train travels down the track: it will rev high during acceleration and then die down once cruising speed is reached: it will rev up again when a grade is approached: when you set your car on cruise control, you will still hear fluctuations in engine speed and notice tach readings change if you have a tach, the automobile will need more power and gas when load conditions increase as a grade and the cruise feature compensates by varying throttle position accordingly to keep set speed steady, you'll hear the engine die down again once the car is over the summit
5. unrealistic control of switch levers: control of a switch should be on a first-come-first-serve basis: AI should not snatch the switch lever that I had control of first out of my hands as a manual driver: the only way I can take back control of a switch is tell the AI train waiting at it to STOP TRAIN. Upon issuing STOP TRAIN command, the AI train holding at the junction releases the lever back to me.


What many here may desire is an AI feature with built-in real-world physics calculations. The genius of software lies in talented human software engineers and developers.
 
Last edited:
I only use the AI drivers if I want a good laugh, - though they are useful for taking a train over a distance where there's just plain track and nothing else for them to get tangled up in. Some of the necessary slow down effects can be implemented by using invisible speed markers spaced out in such a way as to make the robot driver's braking a bit more real world and believable. The railways I build are all single track and are based on UK minor lines so I guess while this method works fine for me, but might not for a US line.
 
Some users get far too carried away with their expectations of the AI system. In a real railroad do the drivers operate the switches and signals as well as drive the trains, then load and unload them? Do the drivers decide the path to take and manage the movements of the other trains?

The AI in Trainz is expected to be a driver (of every type of train and under every type of condition), guard, signaller, switchman, dispatcher and loadmaster. If they cannot (yet) get computers to perform all those roles in a real railroad then why should it be expected in a game with the same realism that some users seem to expect?

Yes, it would be nice to have a shortline virtual railroad with 100+ consists all running everywhere without any human intervention but, seriously, what is the fun in that? I bought into Trainz because it lets me, the person who paid for it, drive the trains. Seriously folks, has our hobby become like the modern day child's train set that runs itself and the child becomes nothing more than a bystander? Are we afraid to "get our hands dirty" and solve traffic problems by manually throwing switches and making our own decisions?

Rant over. I am with KotangaGirl on this one.
 
In BART (Bay Area Rapid Transit) system of San Francisco, California, the computers were able to precisely control train braking, acceleration, switching, routing, stopping at stations and door opening for passengers as early as the year 1971. Computer game simulation should also be able precisely control braking, acceleration and stopping of vehicle images in 2019 by calculating distances, train weight, velocities and so forth. I suspect New York and Chicago subway systems have had such autonomous operation as well. In the train yard of Roseville, California, unmanned d/e locomotives running around the yard there are common and have been there as early as 2011.

Yes, this type of software engineering is possible in Trainz if you want to pay extra as a consumer for a future copy of Trainz that incorporates all these realisms but the N3V software engineering department might be on payroll budget.

We should expect to see autonomous road vehicles with people figures inside them and autonomous boats that laid themselves onto trailers at boat ramps in future editions too. How about diesel road tractor-trailers that autonomously back up to drop and hook trailers?
 
Last edited:
There are different users with different priorities.

I bought into Trainz because it lets me, the person who paid for it, drive the trains.
I have rarely driven a train. I tried hauling a 30 car freight, connecting it to another 30 cars and driving along. Frankly, if I want to "drive" something I'll crank up FSX or go to the PS and drive around Monaco for 30 fast laps.

Seriously folks, has our hobby become like the modern day child's train set that runs itself and the child becomes nothing more than a bystander? Are we afraid to "get our hands dirty" and solve traffic problems by manually throwing switches and making our own decisions?

Seriously, folks, it's much more challenging and rewarding (to me) to design and build a section of the world; dress it up with rivers and roads and trees and bushes and cars and people; create industries and ports that require products and create other products; lay track to service these industries and ports; put together trains to carry the freight and passengers; put appropriate signals on various sections of the track; build and program Interlocking Towers; and, most important and challenging, program the AI to make 10, 15, 20 trains all run smoothly.


That's the beauty of Trainz - something for everyone.

Having said that, it's clear the bias of N3V is more toward people like pware as compared to me. That's probably one of the reasons the AI is no better than it is. That's OK. There's enough here to keep me occupied for a long time.
 
There is a big distinction between "driverless" and "computer controlled". The term "driverless train" is often, understandably, taken to mean "no driver" when in fact there is a human driver, just not in the train itself. The "driverless" ore trains, some with over 200 wagons, that are now running in the NW desert region of Australia are actually driven by human "train operators" located about 1,500km further south. While the "remote" human still makes the decisions and sets the control parameters, computers control the braking and many other features of the train operation, just like they do on modern trains with a human driver on board or on aircraft with computers controlling the engine management and flight control systems with a human pilot still sitting in the front seat. The only true computer controlled trains (i.e. no human intervention at all) that are in operation are like horizontal elevators - they run a shuttle service on the one track between terminal stations with no other trains on the track. Some airports have them to take passengers between the terminals.

I have no doubt that since no real trains are involved and that public safety is not at risk, then it would be possible to have a much improved, even at a "realistic" level, AI system in Trainz but at what cost? I would much prefer that the time and funding be directed towards making other far more urgent improvements and fixes. Its a bit like the user who posted a complaint that the stars and the moon are not "realistic" (there actually is no moon and that was one of his complaints) and wanted that fixed as a priority at a time when there were all sorts of other serious issues (and still are).
 
There are different users with different priorities....That's the beauty of Trainz - something for everyone.

Totally agree. I also find it more rewarding and challenging to create entire "working" worlds than driving the trains in them, but that said the driving part is often the "relaxing part" - I just don't get the time to do it often enough.
 
Where autonomous operation should provide realism in RR simulation is that it provides traffic on lines that you, the manual operator, must contend with. I can have nine mainline trains running under self-operation (AI) while my hands are controlling the tenth one. AI, for all its unsophistication, still manages traffic that I must contend with.

If you work for UP as an engineer in a cab, you will only be controlling one train on the line, the one your body is occupying, not 100's of others at the same time. AI is supposed to mimic a number of human drivers at the same time because there is only ONE human me at the PC.
 
Last edited:
AI is supposed to mimic a number of human drivers at the same time because there is only ONE human me at the PC.

Where that falls down both in Trainz and in real world railroading is that the best problem solver is still the human. The AI can operate well enough when everything works exactly as expected and it is not overloaded. When something unexpected occurs, such as through inadequate planning or testing (and when is there ever enough planning or testing?), then the AI system fails because it cannot make the necessary decisions itself. In my experience, the Trainz AI works best when given simple tasks to perform in low intensity environments. Most (all?) mainline railroads would not meet that criteria. At some point you, the human, has to "put your hand" into the layout and "move a stuck train", just like in basement model railroads.
 
Driving is certainly relaxing, I start to fall asleep after about 10 mins.
If I have insomnia I just start Trainz and drive, within a short time I am ready for a good nights sleep.
My preference is to build a sytem that is very complicated and close to realistic. I run many AI trains and only get involved when the odd one gets stuck.
I use simple signalling and in yards very little signalling and the trains operate by line of sight.
Each to his own is my motto.
Cheers,
Mike
 
I don't know who ever told you that AI would be realistic and perfect, at all … But whoever told you that … deceived you, or you are deceiving yourself, hoping for AI perfection … Perfection and realism are 2 things that you will NEVER, EVER, attain in Trainz … It's just a GAME, man
 
Last edited:
Where that falls down both in Trainz and in real world railroading is that the best problem solver is still the human. The AI can operate well enough when everything works exactly as expected and it is not overloaded. When something unexpected occurs, such as through inadequate planning or testing (and when is there ever enough planning or testing?), then the AI system fails because it cannot make the necessary decisions itself. In my experience, the Trainz AI works best when given simple tasks to perform in low intensity environments. Most (all?) mainline railroads would not meet that criteria. At some point you, the human, has to "put your hand" into the layout and "move a stuck train", just like in basement model railroads.

This sums it up.

My Gloucester Terminal Electric route - my knock-off and highly modified version of George Fisher's Gloucester Terminal, has about 18 active AI driven consists with a dozen or so tram drivers busily handling the X-shaped route on near-endless circuits. The other six or so consists get introduced as outside connections between Boston and Rockport. Since mainline railroad is double tracked, there is no worry about AI getting into a fit and pulling a stupid move. They simply go to and from Boston and Rockport and vice versa. This traffic is a simple mix of passenger and freight trains and are there to add some interest up and around that side of the otherwise busy route with the AI arriving from the two portals at 15 minute intervals. The remaining tram drivers, mentioned, make their loops a dozen or so times then eventually return to the car barns to wait an hour then go out and do another dozen or so routes before repeating.

I developed this route about 10 or so years ago now, and it has taken about 6 years to get the AI to cooperate now to a point where I no longer have to worry about them for most of the time. Using the junction controller assets, I've worked out loops and crossovers so that I no longer have AI driving through each other about 90% of the time. My signaling took some tweaking and planning, and lots of time investment to get that working as intended, and all of this became what I consider to be the biggest challenge in most of the session development.

While the AI is running, I will switch industries, tie up the "mainline" with a switcher and a few boxcars, and sometimes a more formidable freight I bring down to a bigger terminal at the other end of the route. This adds to the mix and challenges and can make for a nice operating session when it works well. For some reason I noticed that this sends the system into a tizzy and everything falls apart rather easy.

Now even with automatic system running at 90-95% success rate, not counting me sticking in a freight train, I still have to stick my hands in there a la old-fashioned model railroading and untie up some stupid drivers who some how take a branch not meant for them, or head down a siding and get stuck. It happens as I've noticed as time wears on and the AI get "tired". The program scripts must bog things down to a point where signals no longer respond quickly, AI start skipping stations, and everything begins to fall apart. By this time I've had it with driving too and I usually quit myself.
 
One thing that I have learned in my career in IT is "never say never"

Exactly and this is why getting something to work in Trainz goes beyond throwing the route or session in the trash and walking away. That career for me was both challenging and fun at times.
 
Where that falls down both in Trainz and in real world railroading is that the best problem solver is still the human.

And yet, in the most complex of all scenarios, our roads, driverless cars are quickly becoming a reality.
 
Last edited:
I have no doubt that since no real trains are involved and that public safety is not at risk, then it would be possible to have a much improved, even at a "realistic" level, AI system in Trainz but at what cost? I would much prefer that the time and funding be directed towards making other far more urgent improvements and fixes. Its a bit like the user who posted a complaint that the stars and the moon are not "realistic" (there actually is no moon and that was one of his complaints) and wanted that fixed as a priority at a time when there were all sorts of other serious issues (and still are).

Comparing N3V modifying its software to fulfill the desires of one person wanting a moon to fulfilling the wishes of hundreds of posters, some even in this thread, asking for an improved AI is really a stretch.

Many of us "would much prefer that the time and funding be directed towards making ... improvements and fixes" to the AI.
 
Many of us "would much prefer that the time and funding be directed towards making ... improvements and fixes" to the AI.

As you stated in post #5, different people have different priorities. One group will demand that they fix "X" as a matter of high priority, a second group will insist that "Y" is far more important while a third group will demand that "X" and "Y" be left exactly as they are. Who would want to be a game developer?
 
Back
Top