I'm getting around 75% prediction accuracy (500 row moving average) on one CA column. Strangely, this doesn't seem to change much whether the predicted column or adjacent columns are being encoded. Swarming doesn't always give me encodings of the predicted field. Sometimes it encodes adjacent fields, but with the same general prediction accuracies, which makes some sense because in CA the history of adjacent rows are causal factors in the current state of the cell.
I even manually updated the model params to encode every single input field, but with very similar prediction results (and a slower runtime, of course). Interesting results. I think there are manual tweaks that could be made to improve accuracy, but I'm pretty happy to predict a pseudo-random pattern with 75% accuracy! https://github.com/nupic-community/nupic.ca --------- Matt Taylor OS Community Flag-Bearer Numenta On Sun, Jan 18, 2015 at 2:18 AM, Fergal Byrne <[email protected]> wrote: > Hi Matthew, > > This would be a great demo (Wolfram's CA stuff appeals to most of us nerds). > I predict that if you feed a fixed set of bits into NuPIC, the TM will learn > the rule you've picked and will be able to predict the next pattern for all > but the edge bits (which will be partly random as far as it can tell). I'd > also predict that a single-order TM (one cell per column) will be also able > to do this learning. > > These two predictions come directly from the CLA theory (Subutai can verify > this), so it could be a good integration test for new implementations > (assuming NuPIC matches my predictions, of course!). > > > Regards, > > Fergal Byrne > > On Sat, Jan 17, 2015 at 10:23 PM, Jeff Fohl <[email protected]> wrote: >> >> I used to be a bit of a cellular automata nerd. I would be interested in >> seeing what you discover. You could also possibly just feed in the values >> for the center column of rule 30 - though that has been shown to be highly >> random, so I am not sure what the utility of it would be? >> >> - Jeff >> >> On Sat, Jan 17, 2015 at 1:59 PM, Matthew Taylor <[email protected]> wrote: >>> >>> I've always been fascinated by elementary cellular automata [1]. Some >>> rules produce interesting pseudo-random patterns with repeating >>> features. I think it would be interesting to see if NuPIC can decipher >>> these features from the randomly generated output of the automaton and >>> predict the continuation of partially-developed features. I also >>> wonder what the anomaly scores would say after NuPIC has seen several >>> thousand rows of data. >>> >>> I've put together a *very* simple program [2] to generate the output >>> of Rule 30 [3], but I did it in JavaScript out of habit. I really need >>> it implemented in Python to get decent integration with NuPIC. >>> >>> To feed cellular automaton data into NuPIC, I assume I'll need to >>> choose some number of adjacent columns within the automatons' output >>> (maybe 10 fields?). Each field would be simply binary, and I've got >>> some code in place now that can extract the columns and print them to >>> the console [4]. >>> >>> Is anyone else interested in this crackpot idea? I have no idea what >>> any applications might be, I'm just fiddling around. Let me know if >>> you're interested and we can discuss. >>> >>> [1] http://mathworld.wolfram.com/ElementaryCellularAutomaton.html >>> [2] https://github.com/rhyolight/cellular-automata-engine >>> [3] http://en.wikipedia.org/wiki/Rule_30 >>> [4] http://youtu.be/TT2-aXrmJ6k >>> >>> Regards, >>> --------- >>> Matt Taylor >>> OS Community Flag-Bearer >>> Numenta >>> >> > > > > -- > > Fergal Byrne, Brenter IT > > http://inbits.com - Better Living through Thoughtful Technology > http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne > > Founder of Clortex: HTM in Clojure - > https://github.com/nupic-community/clortex > > Author, Real Machine Intelligence with Clortex and NuPIC > Read for free or buy the book at https://leanpub.com/realsmartmachines > > Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014: > http://euroclojure.com/2014/ > and at LambdaJam Chicago, July 2014: http://www.lambdajam.com > > e:[email protected] t:+353 83 4214179 > Join the quest for Machine Intelligence at http://numenta.org > Formerly of Adnet [email protected] http://www.adnet.ie
