This is something some of us were jamming on during the hackathon event. I've been chewing on this idea since, and have been hung up on how I'd setup the experiments, but didn't even consider the idea of self organization!
On Wed, Sep 11, 2013 at 7:42 AM, Rik <[email protected]> wrote: > Hi, > > I would like to check the NuPIC community's interest in researching the > use of HTM for 'intelligent' networking -- self-organizing networks that > autonomously learn and optimize routing behavior by analyzing and > predicting traffic flow patterns and as such are able adapt to > fast-changing topologies. This is both on the physical/data layer > (LANs/SDN, mobile ad-hoc, IP/internet scale) as well as the > application/content layer, e.g. peer-to-peer/overlay networks, CDNs, more > speculative things like multi-agent systems. > > In the presence of a feedback mechanism at the routing level, routers > could inspect properties of the data being routed and correlate it with the > success or failure of routing decisions -- i.e. whether and which way to > forward data -- and as such optimize the routing behavior on an ongoing > basis, adapting to changing usage patterns of where data originates and > where it is consumed. As such existing networks can be improved for higher > throughput on the backbone and higher 'SNR' at the endpoints. An assembly > of such learning routers could even self-organize a network from scratch > starting with gossip and opportunistic forwarding and building from there. > > In particular, the idea of content-centric networking -- a hypothetical > general-purpose internet-scale store-and-forward mechanism, promoted > here<http://www.youtube.com/watch?v=oCZMoY3q2uM>, > that seeks to logically detach content items from any particular physical > location and instead constantly rearrange their spread over the 'net for > optimal reach by an ever-fluctuating set of consumers and producers -- may > turn out to be feasible with HTM-like learning mechanisms. > > As an experiment, HTMs located at each network node could build and evolve > a world model of all content passing through the node by forming SDR > representations of its atomic constituents and higher-level beliefs about > the nature of the content seen. Some beliefs would overlap with those of > neighboring HTM-based nodes, making them a preferred choice of forwarding > newly arriving content of the same nature that way. What large-scale > phenomena would emerge from this? > > Jeff mentioned "uber-HTMs" in some of his talks without going into > specifics -- might this be along those lines? > > Cheers > > Rik > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
