Under rating threshold (hideshow)
"so could you just learn to maneuvre through the course?" "yeah sure I'll not run into any walls now." *proceeds to clip the easiest corner for the next 400 generations*
Under rating threshold (hideshow)
Hello! Could you add a button to reset AI progress, but not upgrades? Because bikes get dumb and forget how to complete laps.
Under rating threshold (hideshow)
Would be nice if at least one motorcycle maintained the exact same route as the last top score-er so that it would be guaranteed that the motorcycles would only progress positively instead of the best having a worse score than the last round.
Under rating threshold (hideshow)
The best of any given generation should not have any changes whatsoever for the next attempt. As it is, the allowed changes to the best still will set players back a significant margin on some of the maps (particularly #6 which seems to have one particular turn which confuses the hell out of the algorithms)
Under rating threshold (hideshow)
Best bike decided to just slam right into wall after certain turn, and I had to watch 20 generations copy him before first one even attempted to take that turn.
Under rating threshold (hideshow)
I'm excited to see if people will make more games like this one, and to see you improve the learning algorithm! Using learning algorithms as part of idle games is such a great idea!
Under rating threshold (hideshow)
the game needs a "ghost" implemented that stores the best performing setup for each level and when you switch between levels it wont take so long to relearn what worked well earlier.
Under rating threshold (hideshow)
My bikes learned how to skid sideways all the way through the first thin section of stage 9. They still haven't learned how to recover once they're out.
Under rating threshold (hideshow)
You really should have it so that the winner of the last race exists unchanged in the next race, have that as a baseline
Under rating threshold (hideshow)
there are multiple types of machine learning, one being that there are good reinforcement, and bad reinforcement, by creating zones where the motorcycle would try to steer into and steer out of
Under rating threshold (hideshow)
Nice idea and a nifty little game! Would it be possible to be able to tune the randomization amount the same way you can tune the sensors? It's getting frustrating to watch the algorithm be good enough to do 3 laps, but then having the random modifications mean that out of the max amount, 1-3 make it past the first half lap as the others just make a 90 degree turn straight into a wall, time after time after time. I'd like to see the time improve, but roughly 500 generations has lead to an improvement of 0.2 seconds and very little money gain due to this.
Under rating threshold (hideshow)
A variable that can change the mutation rate of the program would help a lot. Also include a recommendation of what to set that mutation rate. It seems to get stuck doing the same thing and adding a mutation would change that much faster.
Under rating threshold (hideshow)
My bikers keep crashing into the same wall for several generations. If i change the points, nothing happens. They just keep crashing into the same wall. On level 2.
Under rating threshold (hideshow)
Fun game, I really like the idea, but have a bit of feedback.
After retirement and currently on 50th gen on level 1, the MCs keep making the same mistakes. I have at least 4 of them crashing in the first bend every single time. The algoritm should consider learning from mistakes (or at least not repeating them) instead of just looking at the longest path.
Under rating threshold (hideshow)
ftw you complete level 3 with gen 72 on your first run, retire, and can't get a single car to do a lap in map1, 180 gens deep in, with a "smarter IA"
Really cool idea but the learning process is a complete failure, entirely random
Under rating threshold (hideshow)
Dunno if this has been suggested, but if you always keep 1 car static so that it does the same as the best from the last generation, dumbing can be avoided because if the new generation screws up then the best will still be from the generation before.
Under rating threshold (hideshow)
It would be cool if there was an option to specify which kind of learning was being used. This appears to be iterative random evolution with a focus on what worked best - but utterly agnostic of what works the worst. It would be neat to see how different neural networks would tackle the problem, taking both what worked well, and what worked poorly into account for future iterations.
Under rating threshold (hideshow)
i completed level 2 went to 3 saw my bots crash after the U-turn after seeing it happen 3 or 4 times i want back to level 2 and my bots didn't get past the 2 vertical parts in the middle for 40 generations straight all crashing in the same wall ... it's a shame the "learning" aspect is dissapoiting but overall a good idea for a game
Under rating threshold (hideshow)
Change should be focused near the Crash of the bike that traveled the farthest. This did the best until here. Change a decision around this point. the window of correction should be in seconds not the entire trip. Unless there is a speed upgrade change.
Under rating threshold (hideshow)
Just an idea, perhaps your AI getting stuck issue is related to AI not being incentivized in small increments to get past the hurdle? For example if an AI is making very small increments of progress toward the goal, you may want to make sure that any progress is incentivized, or perhaps even increase creativity incentive if roadblocks are encountered.
Under rating threshold (hideshow)
I think a good way to improve the AI would be to not let it re-do something that it has already failed with. If it learned that veering into the wall at the start is a bad idea, don't let it do it again with one of the random adjustments next wave.
Under rating threshold (hideshow)
Honestly I'm at 400 generations and I have to say I don't think you can call this machine learning. They never learn, they do the same dumbass move time and time again that kills them. Hell some of them still suicide into the wall at the very start, I think 400 generations should be sufficient for them to figure out not to do the thing that kills you 2 seconds into the round.
Under rating threshold (hideshow)
there seems to be a very sudden switch between "i'll just make a slow curve in a random direction" to "oh, this is a corner!"
Under rating threshold (hideshow)
They NEVER learned to speed up or down. They always ran at the slowest possible speed, only learned to steer smarter.
Under rating threshold (hideshow)
I think i'm caught in a low energy state where they turn left after the two 90 degree turns in the third level and can't make a large enough change to get out
Under rating threshold (hideshow)
Yes. could use a counter. No, turning left into the wall at any time is not "slight", or "learned" - it's 'programmed' to do that.
Under rating threshold (hideshow)
Modifications are too small. At level 3 they decided to turn left instead of right and after 70 generations of pretty much no income I decided that it's not fun anymore.
Under rating threshold (hideshow)
I am frustrated that levels 8 and 9 can't be beat on any min speed above 3... and why can't they learn to drift ~~ I need some Initial D here
Under rating threshold (hideshow)
Okay so found a bug.
If you havent beaten level 3 yet but you want to retire because you made mistakes buying skill points the game says you wont gain benefits yet.
But after the retire I still received increased multiplier on money.
Also, an option to completely reset the game would be great.
Now I have max retire rewards without ever beating level 2 but I cant get it back to a fresh run now.