Streams (FRP) library

[ngoodman]

We should implement streams in WebPPL. More generally, the functional reactive programming model would be pretty natural and useful to adopt. If our data were a stream we could do something like data.map(function(d){factor(obs()==d)}) that would extend the execution when data comes in (in this case hitting a new factor). Presumably the return value of marginalization would then be a stream of marginal ERPs.

This would be useful in real-world interaction (see #74) as well as for smart webpages. This would also make various incremental inference models more natural. For instance, incremental parsing would amount to a streaming parse that takes a word stream in.

Use cases:

• Updating inference as data comes in. Here you want to do some more inference to update your marginal when new data comes in. For instance, when new data comes in, a new factor is hit, which updates particles. Or, when new data comes in you extend the state and do more MH.
• Idle vs stream-update: since streams might never end, what should WebPPL do with its "idle time" while waiting for more events? In SMC we could do MH rejuvenation steps while idling, and then particle update and resample steps when data comes in.
• Learning parameters over a data stream. Here you want to do something like a (stochastic) gradient step on designated parameters whenever new data comes in.

With streams in place, it should be straightforward to write glue to tie WebPPL to various event- and stream-driven frameworks such as react.js, RxJS, d3.js. This would make it much more natural to use WebPPL embedded in various applications.

[ngoodman]

Streams may also be a natural setting for amortized inference: learning to update faster / better on new events.

[stuhlmueller]

Idle vs stream-update: since streams might never end, what should WebPPL do with its "idle time" while waiting for more events?

Another possibility is speculative execution: Use the predictive distribution over the next data point to compute what would happen if you updated on some possible value, so that we can update very quickly when the data point actually comes in.

[iffsid]

This seems useful: http://highlandjs.org/