Monty.
Copy Markdown
View Source
Sends a point to the learning modules only when something about it has changed enough to be worth learning.
Mirrors tbp/monty/frameworks/models/sensor_modules.py::FeatureChangeFilter.
Without it a surface agent would hand its learning module a new point on every
timestep even when it had barely moved, which is a different experiment: the
agent would fill a graph with near-identical points instead of sampling the
surface.
The tolerances are the config's delta_thresholds, by feature name, and three
names are not features at all:
n_steps— send anyway after this many steps without sending;distance— send if the point moved further than this from the last one sent;hsv— hue is a circle, so the change is the shorter way round, and the tolerance is one number per channel.
A percept already flagged as uninteresting — off the object, or with a surface the fit could not define — is passed on as it is, as Monty passes it: the filter decides between features and a location only, not whether to send.
Summary
Functions
The filter, and the memory a new episode starts with.
Types
@type memory() :: %{ last_percept: Monty.Cmp.Message.t() | nil, steps_since_sent: non_neg_integer() }
What the filter remembers: the last percept it sent on, and how long ago.
@type t() :: %Monty.Frameworks.Models.SensorModules.FeatureChangeFilter{ delta_thresholds: map() }