Distributed Planning of Distributed Sampling/Experimentation

[josh]

For example: A function calls sub-routines with a given region of hypersurface to explore, feature/value maps of global/local features of various sorts – the routine can then work, add to the map, or call other routines & generate sub-problems. In general it can exit or normally or be cancelled (cooperatively??). The space of possibilities is infinite. And the space of details that might matter to the optimal decision about how to proceed is HUGE too. The focus should be on good engineering supporting the models that are expected to be most interesting & the fine tunings that are expected to be most relevant, while allowing for arbitrary challenge/innovation that might turn out to matter a lot in some cases. The testbed should be capable of saying informative things about how often it matters & what the optimal feature selection for that is.

[josh]

Planning in advance – which software parts would work for both temporal & frequency domain audio analyses vs. which would be written? That’s not high dimensional, but I just wanted to raise the topic of spatial coherence at the architecture level.

[josh]

Problem features of interest include:

Is the problem all in compute space or does it involve external real world manipulations?

Both $$$ cost & time delay costs for tests? Which testing regimes can run in parallel & what is the capacity?

Beliefs about Lipschitz conditions & related measures. Do they exist? Is there a firm uppper bound given? Can the upper bound be probabilistically estimated by some form of sampling?

Beliefs which variables form metric subspaces & which are categorical. For the metric ones, which are isotropic vs. estimating influence.

Are there dynamics that change over the time period of data gathering? If so, how is that modeled? Noise in the response vector with drift? Discontinuity inference from data? Discontinuity information from external communications?

Do we sometimes want to use different but similar problem descriptions to set some kind of priors for search? What forms of problem description best support that?

There’s a lot of open paths. Emphasize having what is available working correctly as software on large numbers of representative problems & using ML to map from problem descriptions to most plausible software assembly/configuration.

[josh]

Consider the development of geospatial intelligence activity maps as an interesting application. Sensor readings can include light photos, thermal, sonar, radar, etc. Self-directed learning can be used to predict levels & changes of interesting activity. And the maps can be used to look for featured that correlate with other time series of interest like market basket activity, security issues, service demand, etc.

[Author]

Posts