Indeed, but long running agents out in the field will drift away from the global optimum especially when exposed to changing data.
what happens if agents could gossip to each other about best approaches? Just as humans gossip about the latest and greatest village pump. Which by the way was important for our survival. For example, say new node A receives a new stream dynamically configured. Its fresh and hasn't seen any patterns. Though it is aware of other agents. The other agents could start gossipping to it. Teaching it the needed patterns for survival. It does this by possibly snapshotting higher layer learning patterns then messaging the newbie. The newbie then applies the learning in a GA fashion. Ie it doesn't completely erase its own learning but applies it in such a way that new learning is incorporated. Now this opens up the possibility of running a heavy swarm, the uploading the learned layer onto one node which then propagates throughout the network of nodes. Learning is _not_ done by twiddling configuration but by swapping learning. Android's design is _not_ something I'd emulate. I'm really not interested in the economic model, manufacturing tech lock down approach. I would never emulate android's Google cage. I'm interested in processes that operate in the wild for long periods of time, evolving, improving. I'm looking to disrupt not entrap. Erik Blas <[email protected]> wrote: >On the topic of configuration, I don't think we need to do away with >it. I >mean the way I understand the thinking behind the configuration via >swarming is that you end up with parameters for a model tuned for that >input and use case, analogous to whittling down a bunch of potential >interview candidates to the ones whose brains best match the problem >space >I'm hiring for. > >A small, autonomous unit doesn't need to do the swarming on it's own, >but >could just have the pre-tuned model params embedded and changeable, and >swarming can be done at the manufacturer, or provided as a service to >more >finely tune the embedded software. My android phone doesn't come with >everything necessary to build/package the android os, but I can get new >images from the manufacturer or build my own to enhance the >functionality. > > > >On Mon, Jan 13, 2014 at 9:53 AM, Stewart Mackenzie ><[email protected]>wrote: > >> Long discussions at the dining table with Rik over whiskey on the >rocks >> brought about a few things: >> >> I should wait till the next day to write this but I believe I have an >> understanding audience. >> >> * Configuration: How much of a role does configuration of regions >come >> into play to compensate for no hierarchy? >> >> * Configuration: I'm starting to understand that simplification of >the API >> is one thing but the reduction of the Configuration API or preferably >the >> _elimination_ of configuration entirely is the jackpot. Thus doing >away >> with swarming. Its impractical for a low powered 'internet of things' >unit >> in the field to spawn up a particle swarm optimization to find the >global >> optimum in the face of changing data. we can only find 'our' ('our' = >human >> experience) global optimum through study and experience. >> >> * Hierarchy or lack there of: Just as computer scientists shortcuts >> general AI via classical AI techniques foregoing biomimicry. So are >we >> nupic'ers foregoing proper learning by encoding stuff like days of >week >> into the implementation thus compensation for a lack of hierarchy? >> >> I need to stop now. >> >> >> >> >> >> Kind regards >> Stewart >> >> -- >> Please excuse the typos and brevity >> >> _______________________________________________ >> 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 Kind regards Stewart -- Please excuse my typos and brevity _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
