factor_agent.config.TrialGraphConfig#

class TrialGraphConfig(gamma=0.75, iterations=100, tolerance=1e-08, weights=<factory>, min_jaccard=0.2, structure_neighbours=30, min_correlation=0.1, correlation_neighbours=30, nearest_accepted=10, max_expansions=5)[source]#

Bases: object

The Trial graph’s parameters (_trial_graph), by default CosmosAlpha’s (arXiv:2611.04555, Tables 6 and 7).

With round.picked_parents the run snapshot records every value at the first start, and a resume may not change them.

Attributes#

TrialGraphConfig.correlation_neighbours = 30#: The correlation edges each trial keeps, the strongest.

TrialGraphConfig.gamma = 0.75#: The share of a utility that is propagated in from the related trials, in [0, 1).

TrialGraphConfig.iterations = 100#: The propagation steps at most.

TrialGraphConfig.max_expansions = 5#: Trials built on one trial (its Lineage’s attempts) at which it is not picked again.

TrialGraphConfig.min_correlation = 0.1#: The smallest absolute correlation that makes a correlation edge.

TrialGraphConfig.min_jaccard = 0.2#: The smallest Jaccard similarity of two trials’ subtrees that makes a structure edge.

TrialGraphConfig.nearest_accepted = 10#: The accepted factors a trial’s novelty is measured against, the nearest.

TrialGraphConfig.structure_neighbours = 30#: The structure edges each trial keeps, the strongest.

TrialGraphConfig.tolerance = 1e-08#: The largest move of a utility at which propagation stops.

TrialGraphConfig.weights#: Each relation’s share of the propagation, summing to 1: generation (a trial and its parent), structure (shared subtrees) and correlation.