The following dead code is being removed from src/tbp/monty/frameworks/models/goal_state_generation.py soon. It has some notes sketched down, apparently for computing a confidence metric while taking into account a parent Learning Module.
Although it seems like the right choice to remove the dead code, I wanted to save the notes for later. It was created in the initial commit by @vclay – so attribution is a little unclear (especially with “TODO M”, which may refer to someone with a name starting with “M”).
def _compute_goal_confidence(
self, lm_output_confidence, separation, space_size=1.0, confidence_weighting=0.1
):
"""Calculate the confidence of the goal-state.
The confidence is based on the e.g. separation in hypothesis-space between the
two MLH, and the confidence associated with the MLH classificaiton of the parent
LM. Currently just retuns the confidence of the parent LM but TODO M implement a
more sophisticated function.
TODO M How to normalize the displacement?
Could put through a sigmoid, that is perhaps scaled by the size of the object?
Could divide by e.g. the size of the object to make it likely to be <1, and
then clip it; that way any subtle differences between LMs is likely to be
preserved, i.e. rather than them all clipping to 1.0; can then just make
sure this value is weighted heavily compared to confidence when computing
the overall strenght of the goal-state.
- size of the object could be estimated from the minimum and maximum corners
- or use the max size of the graph --> Note this doesn't account for the
actual size of the object, and these grid-models are not currently used
Returns:
The confidence of the goal-state.
"""
# Provisional implementation:
# squashed_displacement = np.clip(separation / space_size, 0, 1)
# goal_confidence = squashed_displacement + confidence_weighting
# * lm_output_confidence
return lm_output_confidence
