Under rating threshold (hideshow)
its like the algorithm shits itself on later levels. why do they still run into the wall after 2,000 "generacion"s???
Under rating threshold (hideshow)
The word upgrade is misleading, certain 'upgrades' will absolutely destroy the AI, and being unable to reduce a 'upgrade' unless you reset the whole game is horrid, never ever upgrade max speed for example, i saw no reason to upgrade anything other than steering and population, the rest can easily force you to reset the game. Im worried if the developer actually tested the game, a sensational idea but not the best execution, too much trial and error to make the basics, needs a severe quality of life update
Under rating threshold (hideshow)
Game works until a point where it just wont, AI wont learn based on where mistakes exist, instead it gives random inputs until it eventually manage to make it, then it just follows the previous one, sounds kinda alright until you realize the massive conflicts, i got stuck on phase 4 because all the bikers would just take a sudden turn on the bottom left and rush into a wall, they would never learn because they dont try to make changes based on mistake points, i thought it was lacking upgrades so i went back to level 3 and guess what, i couldnt make even the first lap anymore, they would just never ever learn to take the 2nd left, i realize i had upgraded acceleration, something i havent done since the 2nd retire, no i cant progress on the game anymore unless i reset and avoid a upgrade
Under rating threshold (hideshow)
I managed to complete 3 laps on level 9 with two motorcycles but didn't get the medal. Do I have to then retire as well?
Under rating threshold (hideshow)
tbh, this is not really machine learning, it's beam search in disguise. I'd say from the developer's point of view this is probably a decent attempt of implementing a genetic algorithm, though the main problem is the lack of "cross-breeding". The mutation exists, but since it's only done on the single elite genome it behaves more similar to a simple beam search algorithm.
Not sure how others got problem with it not improving, I'd say that it's mostly due to changes in the environment. if the speed or steering changes the properties of the motorcycle changes and has to be re-learnt, max speed is more problematic in that regard, so basically max out steering and minimize speed and then leave it alone
Under rating threshold (hideshow)
This is so dumb, they dont learn at all. Exhibit A:On level 2 out 21 of 25 units crashes by going in a straight line into the wall. Every run is the same. There is no learning here whatsoever. The remaining 4 cruises the entire lap without issue (usually, sometimes its 5, sometimes the green car crashes). This is utter crap. 1 star.
Under rating threshold (hideshow)
I am stuck in an infinite loop where my motorcycles go through track three fine until about the half way point, which then they will just full hold left and crash into the wall, as I just bought an upgrade I have been stuck here for about 30 minutes (at least) waiting to get an additional 20k, any tips?
Under rating threshold (hideshow)
Thank you for making this concept into a game. I'm also curious as others here about how you implemented the learning logic - 2 years ago! Unfortunately on the training side this game falls short and needs an update - it's quite hard to imagine how quickly and irreversibly you can overtrain a model so that it still hits the same corner even after 100 more gens. At the moment acceleration/speed is completely unnecessary. Max'd 5 pop and 8 slow/steering and made it to the small u-turn in level 9 with ease.
Under rating threshold (hideshow)
I wqould suggest that if the game detects the same bike "wins" 10 times in a row but never corsses the finish line, that it automatically throws out the current batch and goes to an older, or even maybe the begining, winning car
Under rating threshold (hideshow)
Could the run be reset when the last blue bike is destroyed? The green bike is just following the best route run so far without any variation, nothing about its route is going to improve matters.
Under rating threshold (hideshow)
look, its a good concept of the game. but trying to implement sharing the game in this state over fixing the game is not the way to go. fix some of the bugs and do some improvements to the way the ai learn and then share the game. It has so much potential for it to just be ignored. Thanks :)
Under rating threshold (hideshow)
too much "random" alteration to courses early on, no where near enough at the important part where you'd actually make progress...at the very least there should be some form of pruning to bad logic in addition to promotion of good logic, like whatever leads to the lowest 10% of progression should be inhibited rather than repeated OVER AND OVER AND OVER I swear I see this crashing a good 1/5 of the bikes within 2 seconds of EVERY new wave
Under rating threshold (hideshow)
I'll be the first one to admit I know nothing about the programming/coding needed to correctly implement machine learning on a game like this. However, I just dont understand why if your green bike (best bike) makes it a certain distance, why dont all other bikes take the same route and then extrapolate from the point of the green bikes death or maybe a little ways before. There is absolutely no reason why bikes should be crashing into the wall 2 feet off the start line once you get a few generations into the track.
Under rating threshold (hideshow)
This game has lot of potential and it deserves more. Better graphics, customizable neural network and better 2D top-view car physics, longer more complex tacks and variety of cars.
Under rating threshold (hideshow)
The main issue I have with this game is that there is no way to look into the "brain" of your motorcycles in order to look for the problem and mutuate them to your advantage
Under rating threshold (hideshow)
These bikes are stuck doing min speed. They don't even try increasing based upon sensor condition. Give us more control over the "intelligence" of the bikes. Both the randomization conditions and the goal at least.
Under rating threshold (hideshow)
Sorry, but it stinks. The bikes ceep crashing into the same spot, again, and again, and again, and again, and again, and again, and again, and again, and again, and again, and again, and again, and again, and again, and again, ....... completely ignoring any change of sensors or parameters, and keeping to learn nothing at all from every try before.
Under rating threshold (hideshow)
the point of machine learning is that when its stuck in a loop of failing repeatedly it TRIES SOMETHING ELSE BASED ON ALL PREVIOUS EXPERIENCES
Under rating threshold (hideshow)
driver 1 :*does the first two laps perfectly and crashes at the last turn
driver 2 :hey can i copy ur homkwork?
driver 1 :sure but make some changes so it aint suspicious
driver 2 :sure mate
also driver 2 :crashes immediately after the start
Me : Excuse me wtf
Under rating threshold (hideshow)
so I've been watching these cars smash into the same corner for a few hours over a few retires... level 4 as others have stated, is a pain in the booty. Y U NO TURN CORRECTLY?!
Under rating threshold (hideshow)
I don't know about anybody else, but the second to last corner on level 4 is IMPOSSIBLE. I don't know what's with it, the a.i. just slams itself on the corner EVERY TIME IT GETS THERE. I don't know if you still develop updates for this, but if you read this, please check this issue out, it feels like it might just be that I didn't have the right build, but I got no clue.
Under rating threshold (hideshow)
Great concept! Love the idea. Would love to see more configurable options, like maybe allow for fewer bikes to make random changes right at the start and crash into the wall in the first second, or to cause some of your bikes to more closely follow the "lead" bike until they get closer to the crash point to speed up how long it takes to get past the last bad mistake.
Under rating threshold (hideshow)
Love the game, though I don't understand why the categories for the skills exist, sometimes I get stuck and think I could make better adjustments if I could trade between the categories.
Under rating threshold (hideshow)
And also a recommendation, since there seems to be an issue with the AI hardlocking itself into less then ideal loops of smashing into the wall, why not have a soft-retire that retain skill points but still resets the AI? Could be less then frustrating than having to retire on level 6 for example and having to grind back up to it.
Under rating threshold (hideshow)
Regardless of the issues presented with riders repeatedly driving to their doom, I respect the effort put into a game like this, and the continuous updates coming out. Thank you for the effort.
Under rating threshold (hideshow)
Damn this game has such a cool concept, too bad it's broken. They keep crashing at the same wall instead of turning right on track 4, like seriously...