Monty.Frameworks.Models.EvidenceMatching.LearningModule (ExTbpMonty.Core v0.1.0)

How much evidence a hypothesis needs before this step is worth spending on it.

Mirrors tbp/monty/frameworks/models/evidence_matching/learning_module.py, whose module-level part is this one decision. The learning module itself is Monty.Frameworks.Models.EvidenceMatching.LearningModule.EvidenceGraphLM.

Why a threshold at all

A step compares the observation against every hypothesis of every object. Most of them are hopeless, and testing them costs the same as testing the good ones. So a hypothesis is only updated if its evidence is above a threshold derived from the best hypothesis so far: near the top of the field is where a step can still change the answer.

A hypothesis below the threshold is not deleted — it keeps exactly the evidence it had. That matters: an object that looks wrong now can still be right later, and Monty's later steps are what make burst sampling worth anything.

The settings

a number — a fixed threshold in evidence units, ignoring the field;
"mean" / "median" — the average or middle of all hypotheses' evidence;
"all" — the lowest evidence of the field, i.e. update everything. This is Monty's default and what the experiments run;
"[n]%" — the best hypothesis minus n per cent of its evidence, so "within 5% of the leader";
"x_percent_threshold" — the same rule using the learning module's x_percent_threshold, which is the setting Monty's configs actually set.

The last two are one rule with two spellings, so both parse to {:percent_of_max, percentage}; the pairing of "x_percent_threshold" with the module's x_percent_threshold happens once, when the config is parsed.

Summary

Types

config()

A threshold setting, parsed: the forms above, with the two percentage spellings already resolved to one.

Functions

evidence_update_threshold(config, max_global_evidence, column)

The evidence a hypothesis must reach to be updated this step.

parse_config(config, x_percent_threshold)

Reads a threshold setting, resolving "x_percent_threshold" against the learning module's own x_percent_threshold.

Types

config()

@type config() :: number() | :mean | :median | :all | {:percent_of_max, float()}

A threshold setting, parsed: the forms above, with the two percentage spellings already resolved to one.

Functions

evidence_update_threshold(config, max_global_evidence, column)

@spec evidence_update_threshold(
  config(),
  number(),
  [number()] | Monty.Frameworks.Models.EvidenceMatching.Hypotheses.t()
) :: float() | nil

The evidence a hypothesis must reach to be updated this step.

With no evidence at all there is nothing to compare against, so the threshold is zero and every hypothesis — there are none — passes it. That case is the first step of an episode for every object in memory.

The column arrives as the numbers it holds or as the hypothesis set that keeps them. The setting the experiments run — "all" — needs only the lowest of them, and a step asks once per graph, so the set is answered from its bytes: decoding six thousand floats to take their minimum was the largest single Elixir cost in a ten-object step.

parse_config(config, x_percent_threshold)

@spec parse_config(term(), number()) ::
  {:ok, config()}
  | {:error,
     Monty.Frameworks.Models.EvidenceMatching.LearningModule.InvalidEvidenceThresholdConfig.t()}

Reads a threshold setting, resolving "x_percent_threshold" against the learning module's own x_percent_threshold.

Parsed once when the learning module is built, because a setting is configuration: if it is wrong, the experiment should not start.