kleinbl00 posted this HTML5 genetic algorithm car program two days ago. It's awfully addictive.
In that post, cgod then suggested using the seed 'hubski' for the world generation.
I've run it over night, and seem to have topped out at 180.52 d:175.78 h:-0.36/8.19m (gen 758).
elizabeth suggested we have a contest, and I think it's a great idea. If only because I want to get pass 180.52m.
Using the seed 'hubski', the person that gets a car the farthest distance by Tuesday Nov 26, 12PM EST (with a screen capture), gets a Hubski t-shirt.
Unless of course, that person is me. :)
Bring it.
UPDATE: The track must be fixed, and your screenshot must include the track graph.
UPDATE 2: With a distance of 1865.85, mike wins the challenge!
I'm just sitting at work, staring at the piles of algorithmic corpses at that stupid little hill muttering "go...go...go!"
tehstone says that cars can fall off the track at 220 meters. I suppose that's what happened to my car, as the track graph does not get longer than shown. Anyone else get more track than I did in the picture? If the track just ends there, then I suppose I've conquered Mt. Hubski, which is really more like Hubski Falls.
With a distance of 1865.85, mike wins the Hubski genetic algorithm car challenge! Obviously, the cliff at 222 meters or so was not expected. However, plummeting off the cliff with speed, mike's car seemed to cover the most distance. mike wins a Hubski t-shirt! Congratulations, mike! p.s. There are probably a number of arguments that could be made for a more sensible competition. However, I would hate to move the goal post after people started putting such effort into it.
I contacted the author of the program and asked about the track ending. His answer follows. So I had a car make it all the way to the end of Hubski Falls, and then over the edge of the world. I don't have a picture of what that car looked like, but the batch it came from had a heavy front wheel, light back wheel, both large, skinny body with a club like head around the front axle. (See my picture and data in another comment). ---
from: Rafael Matsunaga
Hi Michael! Glad you're having fun with the cars! Yeah, the track ends somewhere between 220 and 300 meters, it's actually 300 1-meter long blocks, and due to all the crazy winding towards the end, it ends up shorter than 300 straight meters. After that, it's a free fall to death for the victorious car. :) Thanks for the link! I love seeing where people are coming from!
In response to the question: "Are there limits on mutation size and mutation rate?" The only way this makes sense as a contest of length is to use the same seed word.
1272? I've been letting this thing run since it was posted a few days ago and this score is definitely the best I've done.
#1: 216.23 d:212.45 h:-7.74/7.67m (gen 830) That's my highest score so far. I got past Mt. Hubski a few times (not much). And I seem to have hit a glitchy area at 212ish. Here's my screen cap (note that it's of a replay): Edit: I'm gonna keep trying, as I see some people apparently hit 300-something.
First Run: Trying Rate: 40%, Size: 4%, Clones: 4, Generations: 160. Score: 149.0. Second Run: Trying Rate: 30%, Size: 4%, Clones: 4, Generations: 160. Score: 163.9. Third Run: Trying Rate: 20%, Size: 4%, Clones: 4, Generations: 160. Score: 216.1. Fourth Run: Trying Rate: 10%, Size: 4%, Clones: 4, Generations: 160. Score: 179.4. Fifth Run: Keeping Rate: 20%. Trying Size: 10%, Clones: 4, Generations: 160. Score: 162.2. Sixth Run: Keeping Rate: 20%. Trying Size: 20%, Clones: 4, Generations: 160. Score: 154.3. Seventh Run: Keeping Rate: 20%, Size: 5%. Trying Clones: 9, Generations: 160. Score: 162.1. Eighth Run: Keeping (best combination tried) Rate: 20%, Size: 5%, Clones: 5 ... Generations: 154, Score: 216.6. Ninth Run: Keeping (elements from the best of the rest) Rate: 40%, Size: 10%, Clones: 10 ... Generations: 127, Score: 216.5. Final Run: Trying (halfway between the best two) Rate: 30%, Size: 10%, Clones: 5 ... Generations: 133, Score: 215.5.
The track ends at about 225 yards. After that the cars fall off the edge of the world and rack up an apparently arbitrary number of points for the distance they plummet before the program kills them off and starts a new generation. (The time it takes might be related to how long it takes the other cars to get stuck and die.) This seems like a bug. mk, you might want to consider changing the goal to something like "who can who can travel about 220 yards, or get about 225 points, in the fewest generations?" instead of "who can get the most points?" Because after about 220, the points and the distance given to a car do not seem related to the car itself.
I blogged about this, in Norwegian: http://matematikksenteret.no/content/2421/Genetiske-algoritm...
I'm sure it is brilliant even though I can't read a single word of it.
I think I am ahead.
I guess working in evo-sim does pay off. #1: 388.61 d:384.07 h:-4295.56/7.25m (gen 91)
This is gonna be hard to beat, If it's even possible. What does the track look like after 250? I thought it just ended there. I'm curious as to what your strategy was to get such good performance in so few generations. Anyways, I'm proud of my flock for what they've achieved so far.
Artificial bottlenecks and Cambrian explosions. I went Gould-Eldredge on its ass. but I cropped wrong and will have to do it again. for me the track dropped off somewhere in the 200's but as long as they have gas they fall forward I guess.
I posted a longer writeup over in the original thread from 3 days ago but here's the overview. I started running this overnight with a random seed (didn't notice other people were using "hubski"). The following morning, a car seems to have "won". The track ends here at about 220m and then the car just falls for about 2000 meters before stopping, although there is no indication of why it stopped. Hopefully the seed you all are using doesn't end like this and it was just an issue with my particular seed.
How do you people get over Mt. Hubski? Damn that hill. I ran around 40 generations with some excellent cars but they just couldn't tough it out and get up that hill. All my top scores just ended up being within less than a foot of each other. I'd also like to mention my Macbook pro starts wheezing like it's dying once this program starts to crank so I'm afraid to let it run too long. Think I might need to finally invest in some more memory.
I dunno if you've tried this already, but clicking the "Surprise" button runs the simulation without the display, maybe that'll help your ailing macbook?
It's kind of a moot point because it stops running once I switch browser windows. Is there a way to change that?
I just pulled that tab into it own window and let it be.
I only ever had one car make it up the hill. The others look like so many salmon failing to get upstream. I think you have to get lucky.
I have a competition in me. I want that shirt and I don't know why. 11/21/2013 8:59 AM PST #1: 354.59 d:350.51 h:-4159.06/8.52m (gen 4313) * Also, everyone should probably just update their top scores in the original post and post updates there as well.
I would be curious to see screenshots showing cars which successfully conquer Mt. Hubski. I have run quite a few generations but only had one champion. Here is a video which shows that an advantagously-proportioned car is not enough; it must also get lucky with just the right approach.
I'm having success at a 3% rate and 20% size. My high is 186.65. I'm commenting more for my own memory than anything, and I'll probably edit this as I keep trying.
This was originally posted around a year ago (not the same exact version, but this looks like a fork of the same git repo) and my biggest issue with it is that the ground didn't change- your cars don't evolve to be the best possible cars, they evolve to be the best possible cars for whatever random track you get, meaning that comparisons will be based entirely on whoever got the easier track! Womp womp. Hopefully this version is different! Happy breeding!