|
My friend found this link, which is really cool to look at. It is using genetic algorithm to tune the traits of the car.
The attributes of the car are:
Shapes of the car body
Location of the wheel
Size of the wheel
The body of the car is made up with triangles, and I imagine the algorithm looks for the best combination of triangles of certain sizes.
There are many ways to look for the best combination of these attributes, one silly way is to simply go through all possible combinations. However, this is impractical because there are so many different combinations. What genetic algorithm does is sample a generation of cars, rank them by how far they can travel, and let the high-ranking cars exchange traits (mating) to form the next generation, then repeat. So all in all, genetic algorithm is nothing special or mystical, it is merely a way of finding a good combination of traits in a space of huge number of combinations. So you can call genetic algorithm a kind of "heuristic guided search".
I'll try to explain "mutation rate" here, this is my understanding:
At each generation's end, we're at the verge of developing the next generation. What can we do? Well, we know so far we'd like to take some top-ranking cars, and cross-over their traits. But, this has the danger of getting stuck. For instance, suppose in an unlikely event all of your cars start off identical to each other, then by crossing-over their traits, you actually gets nowhere. What the mutation rate is, then, it means that after the crossing-over, we get a trait for our new car. We then keep 95% of those traits, and replace 5% with something completely random. An intuitive way to think of it is this: You are 50% like your mother and 45% like your father and 5% like yourself, completely independent of your mother and father. Another way to look at it is mutation is the opposite of retention, the more mutation, the more likely you are going to discover new shapes of cars, and more likely you are going to FORGET what has worked well before. A 0% mutation will stop evolution altogether, by not exploring the unknown, whereas a 100% mutation is going to be completely chaotic, by not remembering what it has learned in the past.
I ran the webpage on my computer over 2 days, just to see how far it can get, it seems to hit a plateau of performance after awhile, here's a screenshot:
![[image loading]](http://imgur.com/ygLEJ.png)
Here's the link to the webpage:
http://megaswf.com/serve/102223/
If you follow the first few generations, these car evolve really freaking fast, good "car-like" robots start to form around generation 8 or so for me, alongside with many retard shape cars 
So run the webpage, and shere some screenshots!!!
|
That is so cool, I'd love to see the code of this. But I don't really understand when the "test run" of a car stops. If they fall over it's ok to do a new run but sometimes it starts a new round for no reason.
|
I think it starts a new round when it reaches a new best score or so. This is so cool I can't believe how fast it evolved to a "typical" car shape
|
I was laughing my ass off at some of the terrible cars that came out of this thing, and my dad came in my room wondering what was so funny. Was a bit difficult to explain,
|
On January 24 2011 09:15 gogogadgetflow wrote:
I think it starts a new round when it reaches a new best score or so. This is so cool I can't believe how fast it evolved to a "typical" car shape
There are glitches of that program I think. It sometimes stop the test prematurely. It doesn't stop it when it reaches a new score because logically it's good to give the code more time to run so to attain an even higher score.
But yeah it evolves pretty fast haha
actually take it back. It stops as soon as it reach some high-score, i.e. we know for sure it is the best so far.
|
I feel like I'm watching evolution fail dramatically. lol Still interesting regardless.
|
I've been watching it non stop trying to catch the one baller car that goes like 400, but so far I've seen a lot of unicycles flipping over haha
Edit - yeah like the post under me says, the ones that evolve an arch to prevent getting stuck and a little thing on the back to stop them from flipping back are kicking ass. this is awesome :D
|
The funny thing is that these cars seem to develop a shape that prevents them from falling over.
Edit: Like this one ![[image loading]](http://i.imgur.com/ZUEgv.jpg)
|
the cause of the plateau on my box was a hill with a pinacle at distance 180; the cars needed a short wheelbase to get over the hill but a large wheel base to get there.
|
So, what does the mutation rate actually do?
|
Any idea what the mutation rate does?
|
On January 24 2011 09:36 LittLeD wrote:
So, what does the mutation rate actually do?
it's a measure of how much random changes in the car design are put in between generations. It may even be the setting of how much the settings of each car are 'jumbled up' between the cars in the genetic algorithm.
|
I'd guess it changes how much or how fast the different parts are changed. How intense they mutate, if you will.
|
It doenst always evolve the same way. I have one window where it gets consistent ~175 after 6 generations and in the other one it still does retarded shit in generation 25.
|
|
|
On January 24 2011 09:36 LittLeD wrote:
So, what does the mutation rate actually do?
Probably changes how fast the parameters changes and how wildly they change.
|
On January 24 2011 09:40 betaben wrote:Show nested quote +On January 24 2011 09:36 LittLeD wrote:
So, what does the mutation rate actually do? it's a measure of how much random changes in the car design are put in between generations. It may even be the setting of how much the settings of each car are 'jumbled up' between the cars in the genetic algorithm.
Which means that a higher mutation rate yields more random looking results but gives the evolution a chance to create a car that deviates from the properties of its ancestors.
|
Mine just went from 90+ to less than 10 in about 4 generations. Seems like it began to regress through some bad traits that just kept popping up. I guess my cars needed an ice age to reset things.
Edit: This was on 100% mutation rate.
|
On January 24 2011 09:42 uNiGNoRe wrote:Show nested quote +On January 24 2011 09:40 betaben wrote:On January 24 2011 09:36 LittLeD wrote:
So, what does the mutation rate actually do? it's a measure of how much random changes in the car design are put in between generations. It may even be the setting of how much the settings of each car are 'jumbled up' between the cars in the genetic algorithm. Which means that a higher mutation rate yields more random looking results but gives the evolution a chance to create a car that deviates from the properties of its ancestors.
Jeah, if you set mutation to 0 and you have retard cars, you keep getting retard cars. That also works the other way around. So if you set it higher you have more chance of getting better/worser cars
Note: the chance of getting better cars is a bit higher becouse that's what the game selects on but with a higher mutation rate you also get a greater chance of getting worse cars
|
i do not understand the graphic and color o the car is there a meaning ?
|
tbh if you want the best results just stick with 5% I would guess.
|
This is like watching my PC play QWOP...
|
dam, so many time it just stop the progress even if he can still move T___T
|
The orientation of the suspensions is also a variable.
I think 100% mutation means the car will get ALL new traits, while with 0% it will remain the same. So 50% means it will keep 50% of the shape and get 50% new traits.
|
|
|
On January 24 2011 09:50 quirinus wrote:
The orientation of the suspensions is also a variable.
I think 100% mutation means the car will get ALL new traits, while with 0% it will remain the same. So 50% means it will keep 50% of the shape and get 50% new traits.
Sooooo 1% would be the best mutation then right? It would take long to get to perfection but has a really small chance to regress
Sounds about right. get the one good base form and then minor tweak to perfection
|
On January 24 2011 09:56 Vain wrote:Show nested quote +On January 24 2011 09:50 quirinus wrote:
The orientation of the suspensions is also a variable.
I think 100% mutation means the car will get ALL new traits, while with 0% it will remain the same. So 50% means it will keep 50% of the shape and get 50% new traits. Sooooo 1% would be the best mutation then right? It would take long to get to perfection but has a really small chance to regress Sounds about right. get the one good base form and then minor tweak to perfection
@1:
Not entirely, 1% mutation rate can make you get stuck at a local maximum.
@2:
This is the right idea, but remember you can only remember traits for up to 15 cars, i.e. your memory is bounded by 15, and you completely forget what has worked 2 generations ago.
|
On January 24 2011 09:44 type_d wrote:
Mine just went from 90+ to less than 10 in about 4 generations. Seems like it began to regress through some bad traits that just kept popping up. I guess my cars needed an ice age to reset things.
Edit: This was on 100% mutation rate.
Well then if it is 100% you really get no relationship between the generations. You are just rolling random cars at each generation, completely ignoring the previous one.
|
On January 24 2011 10:07 evanthebouncy! wrote:Show nested quote +On January 24 2011 09:56 Vain wrote:On January 24 2011 09:50 quirinus wrote:
The orientation of the suspensions is also a variable.
I think 100% mutation means the car will get ALL new traits, while with 0% it will remain the same. So 50% means it will keep 50% of the shape and get 50% new traits. Sooooo 1% would be the best mutation then right? It would take long to get to perfection but has a really small chance to regress Sounds about right. get the one good base form and then minor tweak to perfection @1: Not entirely, 1% mutation rate can make you get stuck at a local maximum. @2: This is the right idea, but remember you can only remember traits for up to 15 cars, i.e. your memory is bounded by 15, and you completely forget what has worked 2 generations ago.
So in the end its up to you to evolve the best car.
Set mutation high and screw it back each time when you are certain its a good base form. That way you should end up with a semi-perfect car. This is obviously only the case when you can judge when your car has the necessarily basic traits
Edit: 5-10% seems to be the best range to start with. All my other cars are still nowhere the score of the 5% one
|
This is a lot of fun to watch, lol. Mine has created some surprisingly successful unicycles.
|
On January 24 2011 10:13 Vertig0 wrote:
This is a lot of fun to watch, lol. Mine has created some surprisingly successful unicycles.
Mine's simulating atavism. I had a retardo unicycle with a cannon on the front(actually just a tiny vestigial wheel) get away with breeding in generation 1 because everything else was just as bad and now that same design keeps showing up.
I'm sure it's just a coincidence but it's kind of funny that it keeps happening.
|
Hmmm seems cool but I got bored and closed it.
|
Pretty cool. One gripe I have is how quick it detects that your car has failed. Seems like it'll land on an edge, bounce to a wheel that looks like it might actually move far, but it just detects that it didn't move and deems it unsuccessful.
Or even that it stops at random times later down the line.
|
meh, it seems to be stuck at 300. bored of it now:\
But i will be back to breed my superior car race!
|
The funniest part about this whole thing is when you get a really awfully designed one, and you're like "wow I feel almost sad for this poor deformed thing" then it starts chugging along and going further than you expect it to and you start rooting for it.
|
Cheat engine make it WAYYYYY faster 
|
Mine peaked at around the 6th generation, but then quickly devolved. Doesn't seem to keep many of the positive traits really.
Kinda pissed me off how many cars started more or less upside down, despite looking like a potentially good design.
|
On January 24 2011 09:14 uNiGNoRe wrote:
That is so cool, I'd love to see the code of this. But I don't really understand when the "test run" of a car stops. If they fall over it's ok to do a new run but sometimes it starts a new round for no reason.
for each generation it sets a score goal, and when the car reaches that goal and a little after it resets
|
On January 24 2011 10:37 sushiman wrote:Mine peaked at around the 6th generation, but then quickly devolved. Doesn't seem to keep many of the positive traits really. Kinda pissed me off how many cars started more or less upside down, despite looking like a potentially good design.
Yeah all of mine are sorta unicycles to start out with, just the less front-heavy ones flop down so that the back wheel is on the ground. If one would evolve with the back wheel in a different spot I wouldn't even need these economy-sized wheels
|
starting getting consistency (i.e. 3 successes in a row) in gen. 11, this is the standard bike, which looks backwards
![[image loading]](http://imgur.com/xmxvU.jpg)
I'm going to let it run overnight just to see what the "perfect design" is.
|
On January 24 2011 10:39 Turgid wrote:Show nested quote +On January 24 2011 10:37 sushiman wrote:Mine peaked at around the 6th generation, but then quickly devolved. Doesn't seem to keep many of the positive traits really. Kinda pissed me off how many cars started more or less upside down, despite looking like a potentially good design. Yeah all of mine are sorta unicycles to start out with, just the less front-heavy ones flop down so that the back wheel is on the ground. If one would evolve with the back wheel in a different spot I wouldn't even need these economy-sized wheels
I've had a few runs with this as well. It seems to depend on randomly getting this characteristic in the first generation.
|
I'm getting mostly monster trucks now. Stuck at 189 all the time.
|
I just realized that it won't run if I'm not looking at it, darn.
|
Anyone get past 212 on theirs mine always flips over right there
|
I ran this for hundreds of generations when I first saw the post for this on reddit (which is where it is from, since the general section of TL seems to be a 2 day lag of reddit)
Eventually I just got to this near perfect motorcycle design that could always get to a certain point around 250 score, but that part of the track was just impossible for anything to get over, so it had hit its evolutionary ceiling.
|
Is the testing road generated randomly? My road seems to have a demonic valley around 125 mark that none of my simulated designs have been able to conquer yet. (Generation 13, 6% mutation)
|
lol... I got 360 with 0% mutation...
|
On January 24 2011 10:51 Thratur wrote:
I'm getting mostly monster trucks now. Stuck at 189 all the time.
I get stuck there all the time as well.
My cars mostly have a big frontal wheel and a smaller one at the back. Unfortunately either the frontal one gets stuck at the peak at 190 or the car gets pierced by the peak if the frontal wheel should actually get over it
|
I love this, this is so much fun, thank you for sharing o_o
lol... I got 360 with 0% mutation...
Nice! it seems perfecting your existing traits rather than introducing new ones worked out rather well. I'm going with a 1% mutation myself. I think the 5% you start with is a bit too much because it doesn't give the generations enough time to work out which mutations work well before introducing new ones. Also, 0% is probably optimal for a good number of beginning generations until you start leveling out on the logistic curve of your average distance traveled graph.
|
On January 24 2011 11:01 OpticalShot wrote:
Is the testing road generated randomly? My road seems to have a demonic valley around 125 mark that none of my simulated designs have been able to conquer yet. (Generation 13, 6% mutation)
You're right, it seems like it generates a different track for every person, so comparing scores is not very meaningful.
|
Wow, this is fascinating. I've been watching this for over an hour at 3 in the morning.
My algorithm seems to develop a motorcycle-like archetype: large wheels, long distance between axes, large clearance height, low center of gravity, and no parts sticking out to the front or back.
My other test run develops a relatively tall type of vehicle with a small front wheel and extremely large ground clearance suited for the rough terrain that was generated. The body has an arc shape. It basically runs on stilts so the body never touches the ground.
It's funny how increasing the mutation rate can quickly destroy the progress made by past generations by breeding malformed abominations.
|
What do the red and black graphs mean? I'm assuming them is average distance per generation or something but I can't figure it out.
Also, it seems like the colors have to do with mutation & crossover... can anyone figure out what they mean?
|
On January 24 2011 11:28 Wr3k wrote:
What do the red and black graphs mean? I'm assuming them is average distance per generation or something but I can't figure it out.
Also, it seems like the colors have to do with mutation & crossover... can anyone figure out what they mean?
Generation #0 is all single colored designs, after that the colors show you from where it got the parts used
|
Wish I could show you my everyday study I'm soon done with my master in science studying AI.
This is very simple compared to the heavy lifting AIs out there.
![[image loading]](http://img269.imageshack.us/img269/2541/94216613.png)
There's alot of luck getting the "perfect car". The GA algorithm will very fast find a local optima - when it does that, try setting the mutation to 20% and see it start searching other places in the search space.
|
hmm i tried it for a while never got past that 212 ditch
i might try leaving it on for over a night with low mutation and see how that works
|
my track seems to be very steep at times, and lots of mini-mountains at others, so right away i realized it will favor large wheels that are far apart, with an unusually high 'floor' to be able to move over the mini-mountains without getting stuck
looks like it's getting there:
![[image loading]](http://imgur.com/5rkqt.jpg)
(i turned up mutation rate to 30%)
|
Love it, thanks for posting. I wish that bouncing straight up in the air didn't end the run for a car, has hampered the progress on mine.
1% all the way so far, this bump in my terrain is proving difficult
![[image loading]](http://imgur.com/5bZOG.png)
|
I'm confused, how exactly does it decide when a new generation is going to start?
|
I had one get to 400.11, it wasn't so wide as most of the ones I see in the pictures here. Its wheels were the same size and just about touching, with a triangle tip pointing out the back of the rear wheel to keep if from flipping over.
|
Really big wheels with a low riding shell seem to be the most popular trait of my cars (I'm simulating at a 10% mutation rate, more dramatic failures, but I think it will pay off with respect to my average success per generation).
I may be wrong though. They seem to be getting pretty top heavy.
On January 24 2011 11:01 evanthebouncy! wrote:
lol... I got 360 with 0% mutation...
Creationist?
|
Pretty interesting thing. About the graphs, I think the red is the farthest scorer of a generation. The black might be something like average of a generation.
I've had two running while i played some LoL matches. The first I had at 5% mutation, and it's currently at Gen. 69. The red graph seems to have plateaued, with the black graph slowly trending upwards. So if the graphs do represent farthest and average, that means that each generation is slowly improving overall, but no single one can get past the red line's plateau(which is ~170).
![[image loading]](http://imgur.com/2Zshy.jpg)
The second one running I set up to 25% mutation It's up to Gen. 175 and has made some intersting looking cars. Even though the mutation is still relatively low, it still manages to make several failures you'd expect to see in the earliest generations. The red graph is very inconsistent, with a few peaks of about equal height and a lot of random smaller peaks. But the black graph is somewhat more consistent(by comparison at least.)
![[image loading]](http://imgur.com/4coVG.jpg)
So far it seems like no matter what the cars get limited by the track. Eventually, it generates some nearly-impassable portion. I'll leave them both up overnight and see what happens.
|
Yeah at some point I had it so all twenty were consistently going from 140-180, I left it up all day and no matter what they would never pass the 180 point, it was a really steep incline after a set of two pits (at the 140 point) which would trap a good two thirds of the cars.
|
For my honours thesis, I wrote genetic algorithms to solve the travelling salesman problem. The algorithm that I wrote was modeled on "island mutations". Each "island" had their own local mutations and the best of them would migrate unto other islands. This hopefully helped prevent local-optimums.
|
Just got to generation 20 on default settings (5% mutation rate) and they are starting to adopt useful shapes.
Really fun to watch, thanks for the link.
|
I managed to get to 369 at a 28th generation. It gets harder from there.
|
I'm pretty sure 100% mutation rate will make your cars completely random, like reloading the page for each generation, or even each car if you will.
|
broke 400
![[image loading]](http://oi53.tinypic.com/1znpzyc.jpg)
|
Interesting link, but my cars are peaking at 214.5 (431.3ish) since it drops down from a huge hill into a ditch where the frontal wheel gets stuck.
|
These are what mine are looking like, and they work quite well, the spike keeps them from wheelie-ing, but jagged areas sort of stump them. I'll be interested to see what it develops to conquer that. I love this game.
![[image loading]](http://img248.imageshack.us/img248/2583/ss20110124210102.png)
EDIT: For a few generations it started to develop an annoying habit of getting stuck on the back wheel and stumbling through, it still worked but could have been so much faster. I guess it worked that out, because it's back to normal Dump Trucks now.
It'd be interesting to have the option to design our first car, and let it evolve from there, to see if evolution could make any improvements on what we decided to be perfection.
|
United States17042 Posts
there must be an optimal rate of mutation that's something between 5-15%. High Enough to make sure that you don't settle on a local maxima, but low enough so that you don't throw away the good stuff before it develops.
|
On January 24 2011 18:34 GHOSTCLAW wrote:
there must be an optimal rate of mutation that's something between 5-15%. High Enough to make sure that you don't settle on a local maxima, but low enough so that you don't throw away the good stuff before it develops.
I think the best solution would be an adapting mutation rate: Increase the mutation rate if you cars seem to get stuck at a particular point, decrease as long as they are still making progress.
|
I started the webpage and on Generation 4 and Generation 5 I had some very good trips with a score as high as 150+ like 2-3 times
i like the sit a lot, thanks for sharing :D
|
On generation 15-ish, my cars started developing the perfect shape to go off this hill and spin in the air exactly the number of times required to land on their wheels, past a few nasty ditches. Beautiful.
|
This is really cool reminds me of a computer program Richard Dawkins made in one of his books.
|
![[image loading]](http://i.imgur.com/wLart.png)
the spike in the front wheel prevents it from going to fast an tip over lol
|
So I was casually watching mine and I was pretty far along in generations so my cars were all looking similar, then I had a mutation that had all the normal components except it mutated without a front wheel. Think I woke up my neighbor from that laugh.
|
With 246.5 this was the best car I had so far:
![[image loading]](http://66mb.de/download.php?file=/115342clipboard2.jpg)
|
It would be cool if you could "play god" and make one change per generation to guide evolution in a certain direction, or make a negative change and see how later generations eliminate the weakness.
|
![[image loading]](http://imgur.com/p7TWD.jpg)
Had it working overnight and now there are quite a few looking like real motorcycles. Interesting program, nice find!
|
Damn... I clicked the link while very high...
edit: I can't take it anymore! When will they grow up and stop doing wheelies??
Also, I have a funny one with a very small back wheel, which looks like those penny-farthings :D
edit2: why is there a baby-carriage among my bikes?
this is amazing...
|
![[image loading]](http://i.imgur.com/p1o8Q.jpg)
0% Mutation means very little variation. The tail of the car is useful, because whenever it flips back, instead of going all the way to it's back, it'll land back on it's wheels.
|
gonna let it run overnight with 1% mutation... gonna see what's up
|
It's quite fun to watch, thx.
Unfortunatley the vehicles and especially the score is not really compareable, since it depends so much on the generated track. I for example, just have a track that's really easy in the beginning, then there come some "hills" which are difficult and at one spot there's just a non passable mountain. Don't think any car could manage to overcome this, they just don't have enough power:
Also the 20 samples are a bit too few, in my opinion. In my previous try, after some generations all cars had their wheels fixed to the same triangle, which doesn't allow an edge to the inside between the wheels and thus all cars had problems with sharp tips on the track, when I did it the second time, some looked similar to the one in the above picture and suddenly most of my cars did pretty well.
(Edit: Maybe I should have turned the mutation rate higher.)
|
I had some pretty good bike-like genetic setup going before I went to class, never spawning disabled and reaching nearly 400.
Then I get back home only to find this little number.
![[image loading]](http://imgur.com/jrXeD.jpg)
|
I had a very very good car at one point the back wheel was like a dot but it had insane momentum. The car had a very good baseheight and could climb any cliff. It could have gone very far. If it was not only just the 4th generation and I could only go to 100 lol. IT was very slow but unstopable cause of the moment the backwheel could deliver 
|
This is actually pretty entertaining
|
Ugh, this is getting frustrating. I went backwards and hit a brick wall at generation 10, i'm at 22 now. My car's traits are being destroyed by a few poisonous genes that are dominating. For example 99% of my cars now have a point sticking strait up that makes it top heavy. It took me forever to realize that at the 125m point, where most of my creations fail, it causes the cars to flip, roll over, and gain an extra 3 or 4 extra points. Of course this is terrible for the stability that I need to actually make it over the 125 hill...
Also around generation I0 would get 2-3 vehicles per generation that would make it past the 125m mark. But the the unreliable traits that would let it make it past the 125m hill died off in favor of traits that would reliably make it to 125m, and then die constistantly. :-(.
So far my best run is 227m at generation 12, but the traits that let it happen have long since died out.
|
Is the car allowed to flip over entirely?
I've been waiting to see a car like this, but it never happens.
![[image loading]](http://www.geekalerts.com/u/gbound2.jpg)
|
On January 25 2011 01:14 trainRiderJ wrote:Is the car allowed to flip over entirely? I've been waiting to see a car like this, but it never happens.
If it it flipped it would just start going in the wrong direction
|
On January 24 2011 18:34 GHOSTCLAW wrote:
there must be an optimal rate of mutation that's something between 5-15%. High Enough to make sure that you don't settle on a local maxima, but low enough so that you don't throw away the good stuff before it develops.
Yeah, that's the problem with genetic algorithms, they tend to converge at local maxima.
|
On January 25 2011 01:18 howerpower wrote:Show nested quote +On January 25 2011 01:14 trainRiderJ wrote:Is the car allowed to flip over entirely? I've been waiting to see a car like this, but it never happens. If it it flipped it would just start going in the wrong direction
No it would not...roll a paperwheel, or whatever lies around in your room and try flipping it, so, that it rolls in the other direction. But you mustn't flip it out of it's two dimensions (how to explain...just only flip it the way, it could in the car-simulation).
And btw., I had a car that could flip around, but it happend only about three, or four times and one time, the run stopped, even if it could have gotten further...
Edit:
|
On January 24 2011 18:46 uNiGNoRe wrote:Show nested quote +On January 24 2011 18:34 GHOSTCLAW wrote:
there must be an optimal rate of mutation that's something between 5-15%. High Enough to make sure that you don't settle on a local maxima, but low enough so that you don't throw away the good stuff before it develops. I think the best solution would be an adapting mutation rate: Increase the mutation rate if you cars seem to get stuck at a particular point, decrease as long as they are still making progress.
There is an algorithm called SARSA which basically adapts as it learns and contains a "mutation". It is also adaptable meaning if the course changes, it will adapt quickly.
|
Also, it's really annoying when the simulation gives up early on my cars when they clearly would be able to continue. Happens most often when the car is traveling fast slightly downhill into a up-slope and bounces straight in the air. It SHOULD land and continue but it ends the run while the car is still in the air.
|
On January 25 2011 02:19 trainRiderJ wrote:
Also, it's really annoying when the simulation gives up early on my cars when they clearly would be able to continue. Happens most often when the car is traveling fast slightly downhill into a up-slope and bounces straight in the air. It SHOULD land and continue but it ends the run while the car is still in the air.
its because it "reach" the goal (the right number) and then they just make a new generation
|
What goal? It doesn't create a new generation when that happens. At least I don't think so, I just have it running in the background and look every now and then.
|
well my new track is totally owning this algorithm
there's 2 parts to it:
at about 100, there's a long hill which needs power to work past it. average-sized wheels don't work:
![[image loading]](http://imgur.com/2S8Fg.jpg)
they just can't make it up.
it eventually came up with a design to overcome it:
![[image loading]](http://imgur.com/m38wS.jpg)
I guess the small wheel can generate more force/push and let the bike come up the hill. As a result though, this design has to be front-weighted so as to prevent wheelie-flips from the small wheel on even ground (because the small wheel generates so much power it just flips it backwards)
but then at 130 it runs into a V shaped trap, where a front-weighted car will just flip forward like so:
![[image loading]](http://imgur.com/09vvW.jpg)
soo i think on this track this alg isn't going past 130
maybe i'll be pleasantly surprised but i doubt it
|
I'm on generation 82 with 4% mutation. Even though I think most of them are great this particular track I have has some impossible hills around 160 :/. I'll post screenshots later
|
![[image loading]](http://img213.imageshack.us/img213/4790/bikew.gif)
got to 540 while i was afk
but now most of the time they top out at 200 - 270
|
I declare this the best thing ever. I'm only at gen 10 with a 5% mutation rate and those things are flying down the track!
|
On January 25 2011 01:18 howerpower wrote:Show nested quote +On January 25 2011 01:14 trainRiderJ wrote:Is the car allowed to flip over entirely? I've been waiting to see a car like this, but it never happens. If it it flipped it would just start going in the wrong direction
No it wouldn't. The wheels will turn clockwise no matter how the car flips.
|
Here's my best one so far
![[image loading]](http://imgur.com/PUD4b.png)
|
Best one i got so far (that I saw)
![[image loading]](http://imageupload.org/?di=9129590548616)
|
my best one so far after only 10 generations :D
![[image loading]](http://666kb.com/i/bqe4k8b1lspdotoqt.gif)
|
So I have done like 10 different trials and this is my best so far.
![[image loading]](http://imgur.com/1gWtw.png)
|
so what i thought was impossible (deep pit preceding a long steep hit) was actually beaten, but it seems that it gets harder because those types of cars mess up later on
edit:
lol i just watched one car completely flip over and then continue onward
|
Ah, this remembers me of the year before last year when I took the AI course in my uni.
We had to implement a generic algorithm back then, but none of them was as cool as this lol
![[image loading]](http://img525.imageshack.us/img525/1047/tmpip.png)
What a bike
|
lol i just watched one car completely flip over and then continue onward
I can do one better. I had an entire line of cars who had very rounded, top heavy, roofs because they would flip at the bottom of a very large downhill slope. About 1 in 3 or 4 would do the full flip, just enough to keep the dome top until the trait died off ater 4-5 generations. It was an epic trait.
|
generation 190: 557.4
why is it so damn hard?
|
![[image loading]](http://imgur.com/z19bR.png)
Let it running overnight, doesn't look like it'll ever get up this hill though
|
Heh. Screw your raised-midsection designs. My game found a different solution for those spikes that might catch you.
Fly over them with pure speed.
![[image loading]](http://img254.imageshack.us/img254/8838/namnlsfb.jpg)
|
lol
![[image loading]](http://imgur.com/pbcoK.jpg)
|
what the hell is up with this regression, some bad front-heavy trait took over while I was asleep last night, ruining my sleek designs.
![[image loading]](http://imgur.com/vicBx.jpg)
They all used to look like this:
![[image loading]](http://imgur.com/BL03p.jpg)
|
Would it be possible to add a button to display the best run so far? I noticed that the run cuts off when it reaches the previous best, but still records the rest of the run in the parentheses.
|
United States4126 Posts
I can't believe I watched this run while eating breakfast. I find it so interesting
|
On January 26 2011 05:54 Carefree wrote:
Would it be possible to add a button to display the best run so far? I noticed that the run cuts off when it reaches the previous best, but still records the rest of the run in the parentheses.
I think the parenthesis is the top distance that the current generations cars can get up to, to be considered complete successes, and it's always double of the previous generations top score. When it reaches the number in the parenthesis, it doesn't deem it necessary to keep recording (for some stupid reason) so it restards. Noticed it one time when I restarted, always got confused by that number at first.
|
I've had a simulation running for two days. The track mine spawned has a very large hill at around a distance of 180. It took over 150 generations before a car spawned that could climb it. I wasn't around to witness it though, and it couldn't recreate that success. Yesterday I got home and saw that there were a bunch of generations around 260-280 that built cars that could climb the hill, and luckily enough the first car I watched after I got home made it up successfully. Turns out, the best cars to get up a hill have slim bodies with a small-medium back wheel and a large front wheel.
But, once again it could never recreate it. I left this morning when it was at generation 490 and there's been no further success.
|
This is truly mesmerizing to watch, and it makes me think a lot. How does the selection work? What about mutations? What if you could set different mutation rates for various traits? What if you could have more traits than there are? (Number of vertices, number of wheels, spring coefficients...)
I feel like a version 2 is in order. With various modes of selection, user-made tracks, per-generation random tracks and everything I mentioned above. And blackjack. And hookers.
/goes to read up on Box2D
|
On January 28 2011 08:29 kerr0r wrote:
This is truly mesmerizing to watch, and it makes me think a lot. How does the selection work? What about mutations? What if you could set different mutation rates for various traits? What if you could have more traits than there are? (Number of vertices, number of wheels, spring coefficients...)
I feel like a version 2 is in order. With various modes of selection, user-made tracks, per-generation random tracks and everything I mentioned above. And blackjack. And hookers.
/goes to read up on Box2D
TL should work together to create an advanced version of this tbh. I would help but my programming skills are limited.
|
the cars in that thing dont really seem to be evolving, every once in a while there will be one that goes a long way, and then the next one ends up being complete shit
|
It seems that a new version is out, with the ability to up/down vote the individual cars (?) and some additional options.
http://www.boxcar2d.com/
|
|
|
For those of you who liked this I'm going to do something better. I'm working on a SC2 map that uses a genetic algorithm to distribute points across a variety of skills, such as HP damage range speed etc. Of coure there are going to be a handful of units with differing stats for this to be done with. It's going to be even better as they will have a variety of abilities to use and a variety of behaviors. Stuff such as "retreat after firing", "Use abillity and rretreat after being attacked", "use abillity and attack when the abillity has cooled down", etc.
It gets even better. I also plan on making a way to save the top 50% of your creations to the bank so you can bring them back later and even add a multiplayer option so you can compete with and share genes with other players. It's going to be awesome when I'm finished. I've been working for last 10-11 hours on it.
|
Ok seriously. I have no idea how this works or what its supposed to do or if I was doing something wrong. But It just kinda keeps transforming into piles of crap that never go anywhere. I decided to leave it on for like an hour, and check on it periodically. Never-ending tardfest of non-working cars. What gives? lol
|
On February 07 2011 10:35 arcanejazz wrote:
Ok seriously. I have no idea how this works or what its supposed to do or if I was doing something wrong. But It just kinda keeps transforming into piles of crap that never go anywhere. I decided to leave it on for like an hour, and check on it periodically. Never-ending tardfest of non-working cars. What gives? lol
You probably have a mutation rate that's too high.
|
So I did a little experiment to see how the different mutation rates work. I left 3 different ones running at 0% mutation, 25% mutation, and 100% mutation for 3.5 hours to see what results I would get. I've found that 0% mutation is basically the only way you'd want to go. I think I might check 0% vs other low percentages, but it looks like 0% caps out super fast.
+ Show Spoiler +
|
|
|
Thanks for bump! Loving this, it's brilliant.
|
For people interested in this, there is now a better website about this. Same, but with aditionnal features:
-more control over the algorithm used
-can save cars
-can design cars
-active forum
-...
http://boxcar2d.com/index.html
|
On February 21 2012 16:15 Yenticha wrote:For people interested in this, there is now a better website about this. Same, but with aditionnal features: -more control over the algorithm used -can save cars -can design cars -active forum -... http://boxcar2d.com/index.html
thanks thats excellent
|
|
|
|
|
|
EPT![[image loading]](http://img513.imageshack.us/img513/7779/39203512.png)
![[image loading]](http://img4.imageshack.us/img4/7758/64675835.png)
![[image loading]](http://img24.imageshack.us/i/carny.png/)
![[image loading]](http://img89.imageshack.us/i/boxcar2d.jpg/)
