filter Directory Reference

Directory dependency graph for filter:

Directories
	equalityFilter
	frequencyFilter
	importanceFilter
	similarityFilter

Files
	Filter.cpp
	Filter.h

Detailed Description

Filters are used while writing working-memory snapshots into long-term memory (LTM). They decide whether a new snapshot should be stored or rejected, so LTM keeps useful changes without saving every repeated or low-value observation.

Implemented filters

• Frequency filter: Limits how often snapshots of the same entity are stored. The first snapshot is accepted, and later snapshots are accepted only after the configured time difference has passed according to the best available timestamp.
• Importance filter: Accepts a snapshot when at least one contained instance is important enough. Importance can currently be based on instance confidence or number of accesses, depending on the filter configuration.
• Equality filter: Stores the first snapshot of each entity and then rejects new snapshots that are equal to the last accepted one. Fields with importance 0 are ignored during comparison.
• Similarity filter: Compares a new snapshot with the most recent accepted snapshots of the same entity and rejects it when it is similar enough to one of them. The comparison is importance-weighted and supports primitive values, NDArray, lists, and dictionaries.

Importance values

Importance values describe how relevant a field is for filtering decisions. They can be set in the ARON XML type definition by adding an importance attribute to an ObjectChild field:

<ObjectChild key='position' importance="0.75">
    <float32 default="0.0"/>
</ObjectChild>

The value must be a float between 0 and 1. A value of 0 marks the field as irrelevant for equality and similarity comparisons, while higher values give the field more weight in importance-based comparisons. Not assigning an importance value implicitly assumes that importance is set to 1.