IDyOM Project

h1. Roadmap

A stable release branch (<code>default</code>) is in private beta (will be v1) and will be released under the GPL.  A separate development branch (<code>develop</code>) is also available.

h2. Release branch (v1)

* Fix list of [[Known bugs]]

* Fix built-in examples (Conklin 95 etc.): data doesn't contain all the necessary basic viewpoints.

* Remove absolute pathname in connect-to-database. (mtp-admin/music-data.lisp)

h2. Short-term feature development

Make system more data agnostic:

* Remove dependancy on amuse and mips packages.  (Tested but not pushed.)

* Separation of music viewpoints from model.

* A straightforward language for specifying viewpoints, including viewpoint schemas (e.g. interval, interval size)

Include mtp-admin, ppm-star and amuse-viewpoints code in the idyom reposiory, to allow single download/checkout.

Include unit testing code.

Viewpoint selection:

* Adding viewpoint weights to output.

* Print trace information about VP sets being tested + mean IC values; record this to log file.

* Optionally specify: min-links

* More flexible way for user to specify constraints on viewpoint search:

** Define labelled viewpoint classes

** Pairs/triples of labels/wildcards specify acceptable combinations

** User provides whitelist or blacklist spec

Efficiency:

* Check/extend caching of models etc.

* Use sampling to estimate mean IC during VP selection.

Some basic benchmarks to ensure stability of future development versions.

h2. Mid-term goals

'Pace' viewpoint: measure of information rate (bits/sec), analogous to flow in speech production.

A web service.

Viewpoint selection:

* Optimise based on match with existing IC values.

* Predict more than one basic viewpoint, and provide recommended viewpoints for each one (not just cpitch and bioi). 

Allow user to specify structure of model.

* Determine order in which distributions are combined.

* Specify weights for particular combinations, e.g. weighted viewpoints, or weighted memory stores.

* Multiple memory stores.

* Specify alternative context strategies.

* Provide some prepackaged models, e.g. the current model structure.

Allow models to use predictive information (PI), expected PI and PI rate (as analogs to IC, entropy and entropy rate respectively).

Predict over more than one dataset.

h2. Long-term

Hierarchical structure: chunk common patterns into symbols.

Parallel implementation.