Probably because it is the most efficient way to go
! We may need a way to share already trained AIs with setups!
Probably because it is the most efficient way to go
! We may need a way to share already trained AIs with setups!
im all for prelearning the ai a bit and not let them start from scratch but learning and adaptation to the users driving style are a definite plus
a switch statement isnt intelligence its just mechanic reactions
the point is that learning is the very definition of intelligence something that doesnt learn is far from being intelligent
I second that
Well, at most you could supply better default racing lines, but a part of the learning is tied to the setup, so you can't really provide highly trained AIs that easily. Maybe pre-training them on a special setup (like the RACE_S one) could work, as any training for the "hard track" ones would be wasted effort.
well they shopuld be at least pretrained to the point where theyre aware of the size of the car theyre in, know some basics of driving (a fresh ai will often countersteer way too late in any setup) and have a reasonable racing line in their head
imho its a huge problem that you need to train them for hours until they know their car takes up more 3d space than a dot and to get them to the point where they can find something that resembles the braking point for a corner on any setup
also in this context their mind shound not be entirely split on a by track basis but rather have a central skillset that saves things like car dimensions, countersteering, overtaking ... stuff like that
imho its a huge problem that you need to train them for hours until they know their car takes up more 3d space than a dot and to get them to the point where they can find something that resembles the braking point for a corner on any setup
also in this context their mind shound not be entirely split on a by track basis but rather have a central skillset that saves things like car dimensions, countersteering, overtaking ... stuff like that
How is AI modeled in LFS?
Is it a neural network with back propagation if a lap is better than previous, or genetic algorithm which saves data for best laps and discards bad laps and in the next lap mixes data from 2 laps to try to make a better line
Im really interested in how do they learn?
Is it a neural network with back propagation if a lap is better than previous, or genetic algorithm which saves data for best laps and discards bad laps and in the next lap mixes data from 2 laps to try to make a better line
Im really interested in how do they learn?
No, each time they cross the finish line LFS precalculates a theoretically faster driving line which the AI then takes. The obvious flaw of this is that it doesn't even check if the new line is actually faster than the previous, so mislearning can also happen (like the AI "learning" a line going through the dirt more and more each lap, making them crash more horrible each time). At least it is like that if I correctly remember what Scawen said back then.
E: Actually, search is a fine tool.
E: Actually, search is a fine tool.
Well, technically, any AI will just be mechanic reactions. Computers can't 'learn' like humans can.
I don't agree here, I just love that the AI learn by driving more and more! And hope that Scawen will work with that feature also.
I learn when I drive more and more...so I think it is a nice feature for the AIs also
The "learning system" is somewhat a confusing statement. It really has nothing to do with intelligence.
Sure, they "learn" a better driving line. However, if they are not prone to the same behavioral mistakes a human is, they are nothing but an image which becomes more and more "in focus" until the picture is so clear there is no getting better. I don't know how anyone can call that "intelligence".
True intelligence is the ability to form a solution given a problem with a known set of variables and a set of unknowns (like "is that driver gonna go wide or not"). The "learning system" as I have seen it is nothing like that at all.
The reason the AI are so bad revolves around the learning process. We have all seen what happens as they learn and fuel decreases... only to fill them with a full tank and watch them spin in circles....
Hardly Intelligence.
It isn't completely useless though. I would say it would make a good component IF it reset to a default value at the beginning of each first practice session. So learning resembles going away and then coming back to a track. The default value should represent a skill level of the driving AI. The learning process should be done before hand (by the devs) and contain a set of different degrees of knowledge and behavior which can be assigned randomly to different driving AI at the beginning. The degree of learning should be random also.
Of course we still need the other 90% of the AI.
Sure, they "learn" a better driving line. However, if they are not prone to the same behavioral mistakes a human is, they are nothing but an image which becomes more and more "in focus" until the picture is so clear there is no getting better. I don't know how anyone can call that "intelligence".
True intelligence is the ability to form a solution given a problem with a known set of variables and a set of unknowns (like "is that driver gonna go wide or not"). The "learning system" as I have seen it is nothing like that at all.
The reason the AI are so bad revolves around the learning process. We have all seen what happens as they learn and fuel decreases... only to fill them with a full tank and watch them spin in circles....
Hardly Intelligence.
It isn't completely useless though. I would say it would make a good component IF it reset to a default value at the beginning of each first practice session. So learning resembles going away and then coming back to a track. The default value should represent a skill level of the driving AI. The learning process should be done before hand (by the devs) and contain a set of different degrees of knowledge and behavior which can be assigned randomly to different driving AI at the beginning. The degree of learning should be random also.
Of course we still need the other 90% of the AI.
Viper Racing had the best AI I've seen so far. Quite hilarious as well, as you could scare them off a cliff by swerving a bit towards them without actually having a collision
GTL has some very very good AI.
GTR2 has pretty good AI.
RFactor is dependant on the track and AI files. Some tracks have faboulous AI and some have awful AI. Just depends on the track.
LFS S1 had great AI.
what will the AI become?
(62 posts, started )
FGED GREDG RDFGDR GSFDG