Hey!

The AI in Live for Speed is no priority at all. But for inexperienced racers and newcomers to simracing they offer a first sight into the simulation. At the moment the performance of the AI lacks behind, even on the higher difficulty levels they can be beaten rather easily. One issue is the racing line the AI drivers follow, which is also visible to the regular driver by activating the racing line in-game. To improve the racing lines, I suggest importing the lines from the hot lap charts, i.e. the world records of the official car/track combinations.
That way the AI follows the quickest path around the track and new drivers will be able to have a competitive reference when learning the tracks. It will offer more competitive driving against the AI without the need to touch the current logic of the AI driving each track.
In my naive thoughts this may be a rather efficient way of implementing a more competitive AI system. Alternatives would probably require the complete redesign of the AI logic. Of course if this is a goal in the future, such an implementation would not be of much use. Also if this proves to require more effort then benefit for a in all honesty low priority part of the game, this idea should neither be followed on.

I hope this was not suggested before. My quick search showed me no improvement suggestion, that followed a similar thought.

Thanks and best regards!


PS: Happy new year!

Wow this is an exact question I came here to ask
This morning.

Maybe somehow there is a way to convert the useful information from the RAF files into KNW

I like the idea, but I doubt they'll do that

I think that AI's in lfs use deep neural network algorithm to learn how to drive, but do not quote me on that. In that case *.knw files store the biases and weights for the neuron layers. What I'm trying to say is that they are a able to learn, so you can race against them and teach them how to pass. But as you say, the problem is in the starting point and that is a default path. The good news are that in new patch lfs can generate AI paths automatically. In that process you can see how the AI path generation is itterative and in each step the AI's get better and better track times. Further improvement is possible only if human players is racing against them and they get pretty decent actually, at least for the beginner racer.

So AI's in lfs being a low priorrity is a huge understatement. There has been a lot of progress and developement in that respect already.

I don't think the AI in LFS learn anymore. The original AI did, but like a decade ago Scawen did work on them to make them faster and now they don't learn from racing.

The AI feel really slow even on level 5, I think this would be a good idea so I can learn how to race them without making people angry because my driving sucks

Updating AI trajectories via world records may be a good idea (although I don't think replays of these records are always very realistic). But I don't know if it can be done easily. An AI editor on the model of the sound editor (Shift+A) could perhaps allow everyone to adjust the AI as they wish. Or even more simply to be able to manually adjust the time or speed of the AI.

[I'm asking the question here because I haven't found any recent information and I don't know if it's appropriate to create a new topic on AI]

I don't know if the current AI is still able to learn but it adapts its best times to the lap. (I have several installations of LFS on different computers and AI times are different)

Does anyone know how the LAP time of the AI is defined? And how to do it (Full race? With qualification? ...) for the AI to adapt to the player's best times?

Thank you for your answers

Some of the WR trajectories or at the top positions of the charts are peculiar, to say the least, but the AI could learn a thing or two when negotiating a chicane.

Though I can be fooled, to my knowledge the AI keeps learning the more they drive. To learn the basics before going online I raced against them multiple times, and those of them I used the most (AI 1 to 5/10) seem to outperform other AI elements which I seldom use.

An AI able to perform around 103% of WR would be really useful and fun.