An experimental, and quite angry neural network livecoding synth (with an audio ‘weave’ visualisation) for the ZX Spectrum: source code and TZX file (for emulators). It’s a bit hard to make out in the video, but you can move around the 48 neurons and modify their synapses and trigger levels. There are two clock inputs and the audio output is the purple neuron at the bottom left. It allows recurrent loops as a form of memory, and some quite strange things are possible. The keys are:
w,d: move diagonally north west <-> south east
s,e: move diagonally south west <-> north east
t,y,g,h: toggle incoming synapse connections for the current neuron
space: change the ‘threshold’ of the current neuron (bit shifts left)
This audio should load up on a real ZX Spectrum:
One of the nice things about tech like this is that it’s easily hackable – this is a modification to the video output better explained here, but you can get a standard analogue video signal by connecting the internal feed directly to the plug and detaching the TV signal de-modulator with a tiny bit of soldering. Look at all those discrete components!
I’m working on a top secret project for Sam Aaron of Meta-eX fame involving the Raspberry Pi, and at the same time thinking of my upcoming CodeClub lessons this term – we have a bunch of new Raspberry Pi’s to use and the kids are at the point where they want to move on from Scratch.
This is a screenshot of the same procedural landscape demo previously running on Android/OUYA running on the Raspberry Pi, with mangled texture colours and a cube added via a new livecoding repl:
Based on my previous experiments, this program uses the GPU for the Raspberry Pi (the VideoCore IV bit of the BCM2835). It’s fast, allows compositing on top of whatever else you are running at the time, and you can run it without X windows for more CPU and memory, sounds like a great graphics livecoding GPU to me!
Here’s a close up of the nice dithering on the texture – not sure yet why the colours are so different from the OUYA version, perhaps a dodgy blend mode or a PNG format reading difference:
The code is here (bit of a mess, I’m in the process of cleaning it all up). You can build in the jni folder by calling “scons TARGET=RPI”. This is another attempt – looks like my objects are inside out:
We’ve had tens of thousands of people spotting nightjars and donating a bit of their time to sensory ecology research. The results of this (of course it’s still on-going, along with the new nest spotting game) is a 20Mb database with hundreds of thousands of clicks recorded. One of the things we were interested in was seeing what people were mistaking for the birds – so I had a go at visualising all the clicks over the images (these are all parts of the cropped image – as it really doesn’t compress well):
Then, looking through the results – I saw a strange artefact:
My first thought was that someone had tried cheating with a script, but I can hardly imagine that anyone would go to the bother and it’s only in one image. Perhaps some form of bot or scraping software agent – I thought that browser click automation was done by directly interpreting the web page? Perhaps it’s a fall back for HTML5 canvas elements?
It turns out it’s a single player (playing as a monkey, age 16 to 35 who had played before) – so easy enough to filter away, but in doing that I noticed the click order was not as regular as it looked, and it goes a bit wobbly in the middle:
Open sauces is a FoAM project to investigate the sharing of food, food culture and food systems. Last week in Brussels we started experimenting with ways to store, display and reason about recipes in different ways. Taking the recipes from the Open Sauces book we’re representing them as Petri Nets, which means we can feed them into various different visualisations, from Scheme Bricks – taken from the Naked on Pluto’s gallery installation projection:
To a new brand new circular representation:
These structures are filtered somewhat to be more readable than the raw petri nets, which can be rendered via graphviz for debugging:
Some decent sized screenshots of al jazari and scheme bricks rendered with fluxus’s tiled frame dump command. This set includes some satisfyingly glitchy al jazari shots – not sure what was causing this, I initially assumed it was the orthographic projection, but the same artefacts occurred on the perspective first-person robot views, so it needs further investigation.
Prepare your bicycle clips! Kaffe Matthews and I are starting work on a new Bicycle Opera piece for the city of Porto, I’m working on a new mapping tool and adding some new zone types to the audio system.
While working on a BeagleBoard from one of the bikes used in the Ghent installation of ‘The swamp that was…’, I found (in true Apple/Google style) 4Mb of GPS logs, taken every 10 seconds during the 2 month festival that I forgot to turn off. Being part of a public installation (and therefore reasonably anonymised :) – this is the first 5th of the data, and about all it was possible to plot in high resolution on an online map:
It’s interesting to see the variability of the precision, as well as being able to identify locations and structures that break up the signal (such as the part underneath a large road bridge).
Optimisation is a game where you write more code in order to do less. In Al Jazari 2 doing less means drawing less blocks. Contiguous blocks of the same type are already automatically collapsed into single larger ones with the Octree – but if we can figure out which blocks are completely surrounded by other blocks, we can save more time by not building or drawing them either.
Here is a large sphere – clipped by the world volume, showing a slice through the internal block structure:
The next version has all internal blocks removed, in this case shaving 10% off the total primitives built and drawn:
The gaps in the sphere from the clipping allow us to look inside at how the octree has optimised the structure. The gain is higher in a more normal Minecraft set up with a reasonably flat floor covering a large amount of blocks. Here is the code involved, built on top of a functional library I’m building up on to manipulate this kind of data. It maps over each Octree leaf checking all the blocks it touches on each of its six sides, taking into account that the size of the leaf block may be bigger than one.
(define(octree-check-edge f o pos size)(define(do-x x y)(cond((eq? x -1) #f)((octree-leaf-empty?(octree-ref o (vadd pos (f x y)))) #t)(else(do-x(- x 1) y))))(define(do-y y)(cond((eq? y -1) #f)((do-x size y) #t)(else(do-y(- y 1)))))(do-y size))(define(octree-is-visible? o pos size)(or(octree-check-edge(lambda(x y)(vector size x y)) o pos size)(octree-check-edge(lambda(x y)(vector -1 x y)) o pos size)(octree-check-edge(lambda(x y)(vector x size y)) o pos size)(octree-check-edge(lambda(x y)(vector x -1 y)) o pos size)(octree-check-edge(lambda(x y)(vector x y size)) o pos size)(octree-check-edge(lambda(x y)(vector x y -1)) o pos size)))(define(octree-calc-viz o)(octree-map(lambda(v pos size depth)(octree-leaf(octree-is-visible?
o pos size)(octree-leaf-value v)))
Departing in a new direction after evolved light follower robots, take 500 processor cores spread out in space. Give them a simple instruction set which includes a instruction to copy (DMA) 8 bytes of their code/data to nearby cores (with an error rate of 0.5%). Fill the cores with random junk and set them running. If we graph the bandwidth used (the amount of data transmitted per cycle by the whole system) we get a plot like this:
This explosion in bandwidth use is due to implicit emergence of programs which replicate themselves. There is no fitness function, it’s not a genetic algorithm and there is no guided convergence to a single solution – no ‘telling it what to do’. Programs may arise that cooperate with each other or may exhibit parasitic behaviour as in alife experiments like Tierra, and this could be seen as a kind of self modifying, emergent Amorphous computing – and eventually, perhaps a way of evolving programs in parallel on cheap attiny processor hardware.
In order to replicate, the code needs to push a value onto the stack as the address to start the copy from, and then call the dma instruction to broadcast it. This is one of the 500 cores visualised using Betablocker DS emulator display, the new “up” arrow icon is the dma instruction.
Thinking more physically, the arrangements of the processors in space is a key factor – here are some visualisations. The size of each core increases when it transmits, and it’s colour is set by a hash of the last data transmission – so we can see the rise of communication and the propagation of different strains of successful code.
The source is here and includes the visualisation software which requires fluxus.
Continuing with the structured procrastination R&D project on evolvable hardware, I’m proud to report a pretty decent light following robot – this is a video of the first real-world test, with a program grown from primordial soup chasing me around:
After creating a software model simulation of the robot in the last post, I added some new bytecode instructions for the virtual machine: LEYE and REYE push 1 on the stack if we are detecting light from the left or right photoresistor, zero if it’s dark. LMOT and RMOT pop the top byte of the stack to turn the motors on and off. The strategy for the genetic algorithm’s fitness function is running each 16 byte generated program on the robot for 1000 cycles, moving the robot to a new random location and facing direction 10 times without stopping the program. At the end of each run the position of the robot was compared to the light position, and the distances were averaged as the fitness. Note that we’re not assigning fitness to how fast we get to the light.
This is pretty simple stuff, but it’s still interesting to look at what happens over time in the genetic algorithm. Both motors are running at startup by default, so the first successful programs learn how turn one motor off – otherwise the robot just shoots off and scores really low. So the first generations tend to just go round in circles. Then they start to learn how to plug the eyes in, one by one edging them closer to the goal – then it’s a case of improving the sample rate to improve accuracy, usually by using jmps and optimising the loops.
This is an example of a fairly simple and effective solution, the final generation shown in the animation above:
Some explanation, the right and left eyes are plugged into the left and right motors, which is the essential bit making it work, the ‘nop’s are all values that are not executable. The ‘rmot’ before the ‘jmpz’ makes the robot scan around in circles if there is no light detected (strangely, a case which doesn’t happen in the simulation). The argumant to ‘jmpz’ is 0 (loop) which is actually the 17th byte – so it’s cheekily using memory which has been initialised to zero as part of it’s program.
This is a more complicated and stranger program which evolved after 70 generations with a high fitness, I haven’t worked out what it’s up to yet: