It Takes Two Neurons to Ride a Bicycle (2004) [pdf]
paradise.caltech.edu
paradise.caltech.edu
https://web.archive.org/web/20041204162656/http://www.nips.c...
Virtual Bicycle — Matthew Cook, California Institute of Technology:
http://www.paradise.caltech.edu/cook/papers/index.html
In this work I created an environment in which I assumed learning of higher order concepts would be necessary. I created a general purpose physics-based hinged rigid body simulator and used it to simulate a bicycle for a learning agent to learn to ride. But as others have found in other contexts, most environments do not require higher order concepts, and in this case it turned out that a simple two neuron circuit was already sufficient for controlling the bicycle, indeed, better than a human using the keyboard.
The physics simulator was a significant project in itself: I studied rigid body mechanics (which is more complex than most of us realize) and designed the simulator explicitly so as to simultaneously exactly preserve both angular momentum and kinetic energy, as previous simulations of mine (in quantum mechanics!) had shown me that preserving conserved quantities can be crucial for getting accurate results. The simulator works nicely, and I later read in Sam Buss's 2001 paper, Accurate and Efficient Simulation of Rigid Body Rotations, that my precautions were well warranted.
Later I was able to reduce the controller to just one neuron, and then to an even simpler plain linear feedback system, confirming the finding that many real problems are best solved by a hack. The simulator has been useful in other projects since then.
Virtual Bicycle Download Page:
A discussion from last year: https://news.ycombinator.com/item?id=16215130
Does it have symmetry, yes.
However, this is a simulation and with that the constraints and randomness of the variables will over many runs, highlight if there are any aspects that are overlooked.
Looking at this run, level of wobble seems to be overly bearing with noticeable patterns that for me would question the model.
That and as a kid, we had a grass hill of about 500m in length and a mild graduand and fairly flat that you could ride a bike upon. I would go down it, side straddle my bike, step off and see how far I could run alongside my bike and get back on. Yes there was accidents, but I was a kid, nothing beyond a bruise doing this. I would say being conservative, I easily managed 25 meters. Now this was on a hill, more momentum to start with and the big factor, on grass.
Yet I look at the model runs and see very little running straight down. I would factor in the model has a fixed initial momentum, and a random increasing wobble. This would play out with the pattern being seen.
What's needed now is for this to be done 800 times for real. Even with a static push, factors would come into play. Such as dust build up, temperature changes, humidity can all have a small factor in the outcome.
Which highlights a big area with simulation and real runs comparisons. Simulations will all those nuances chaos factors as static and as if all runs happen at the same point in time. That's hard to repeat for real for comparison. Which makes it hard in nailing all those small areas that make small, but noticeable effects over time.
With that, is the problem a fractal? In that the more you look into all factors that affect it, you end up zooming into a whole new area of factors that affect that factor.
A dickbar at the top telling me to open it in the app.
A small section in the middle of the screen with a 50% black overlay that allows me to see about a quarter of the content with no ability to scroll or close the overlay.
A double wide dickbar on the bottom asking me if I want to open the page in the app or Safari (I use Firefox). I cannot make this go away so I can see the page.
Fuck Reddit.
I'll keep using Reddit from a desktop browser, but I won't install the app. So now they get no revenue from me, when they had the chance to get some.
I was the traditional target demographic of reddit, now I will go out of my way to avoid it. Just like Pinterest
Without an app, anyway. I've been using Diode (available on FDroid, not sure about the Play store) and it's pretty good from what I've seen.
That was a quick uninstall.
1. old, but not a repost should be OK and not really need a date (dates are good when it is a repost or notorious. this is as obscure as it can get)
2. not linking to the original source, but to a insightful discussion instead. I often link to slashdot discussions which contain not one, but 3 or more source on the summary with disparate point of views, plus the slashdot discussion, which is extremely relevant (for hardcore network stuff at least). and the post get deleted by not linking to a original source.
Should I submit the periodic table? Every scientific paper ever posted?
It would waste less time if random posts from a year or older appended the year to the title. Seems very common, but not mentioned in the hacker news guidelines.