Skip to content

algorithm.md

Latest commit

 

History

History
76 lines (55 loc) · 2.76 KB

algorithm.md

File metadata and controls

76 lines (55 loc) · 2.76 KB
Raw

Algorithm and Memory

Compact reference for the system's internal behavior.

Working Memory

Working Memory is volatile and lives in RAM as a NetworkX directed graph. It stores recent auditory episodes, sound nodes, location nodes, activations, recent co-occurrences, focus, arousal, and consolidation state.

An episode represents a concrete sound in a location. If the same sound is heard again in the same location while it is still active, the episode is updated; otherwise, a new one is created. When it has not been heard for more than inactive_gap_s, it is published for consolidation and later removed from Working Memory according to episode_ttl_s.

If no reliable location is available, unknown_location is used internally. That label is not treated as a real room unless explicitly enabled with learn_unknown_location_patterns.

Long-Term Memory

Long-Term Memory is persistent and is saved as JSON at ltm_path. It does not primarily store raw episodes; it uses consolidated episodes to reinforce patterns.

Current patterns:

  • familiarity and count per sound type.
  • sound-location associations.
  • typical sounds per location.
  • sound-sound co-occurrences.
  • approximate time-of-day histograms.

An episode such as 08:05 kitchen: coffee_machine reinforces the familiarity of coffee_machine, its relation to kitchen, and its morning histogram. Repeated routines increase weights and usually reduce the later novelty of equivalent events.

Novelty

Novelty estimates how much an observation violates learned patterns:

novelty =
  0.40 * (1.0 - familiarity)
+ 0.45 * (1.0 - location_congruence)
+ 0.15 * (1.0 - time_expectedness)
  • familiarity: whether the sound has already been learned.
  • location_congruence: whether the sound fits the current location.
  • time_expectedness: whether it usually occurs around the current time.

If location is missing, congruence uses the neutral location_missing_neutral_congruence value instead of treating it as a room mismatch.

Arousal, Focus, and Contextual Urgency

Arousal is a global value from 0.0 to 1.0. It increases with novel, incongruent, or contextually urgent events, and decays over time.

Contextual urgency does not change base novelty. It can make a later sound more relevant when it occurs shortly after a high-novelty event, for example glass_breaking followed by voices in the same area.

Focus is selected from active episodes using novelty, location incongruence, intensity, and contextual urgency. In rqt it is marked with a red border.

Time

The logic uses AuditoryObservation.header.stamp when present. This lets accelerated simulators keep simulated hours consistent for co-occurrence, contextual urgency, and time expectedness. If the timestamp is empty or zero, wall-clock time is used.