liveBook · Manning

Lineage tree

As explained in Self-Improving Agents

A lineage tree is a tree-shaped record of candidate variants and their parent-child ancestry during a search.

When a search stalls, its lineage tree shows where each candidate came from and which branches remain available. In Darwin Gödel Machine (DGM) search, each evolved prompt is an immutable artifact with a pointer to its parent, so the run forms a tree. section 4.4 section 4.3.5

Reading the tree makes the search history useful:

  • Follow a candidate's parent links to trace its ancestry. section 4.4 section 4.4.1
  • Compare branches and archived candidates; keeping them can let the search recover from a stalled main line. section 4.3.5 section 4.6
  • Inspect node scores and parent-relative changes, plus which candidates were available when a champion was bred. section 4.4.1

Recovery from a neglected branch

In one DGM run, the champion grew from a low, structurally different branch that the archive had preserved. The ordinary line had stalled, while a sibling branch produced dead ends; sampling reached back to the neglected branch and continued past the plateau. section 4.3.5 section 4.4.1

Figure 4.14 The Lineage panel tree view over the DGM run, champion selected. Clicking a version lights its ancestry, so the winning line runs genesis to crown while the rest of the archive fades to context, and dashed rings mark the breeding pool at the champion's birth. The champion grew from a branch untouched since the first iteration, its equal-scoring sibling having burned two dead ends in between, and the tree ends at the crown because the run stops the moment ground truth reports a perfect answer.

The Lineage panel highlights a selected candidate's ancestry and shows the surrounding archive for context. section 4.4.1

How Self-Improving Agents covers it

How deep it goes
Exposes parent links, candidate scores, parent-relative changes, and the breeding pool in a visual tree.
Examples
Shows a champion emerging from a preserved low branch after the main line stalled and a sibling branch produced dead ends.
What you can do after
Evaluate a search by tracing a champion's ancestry and comparing it with archived alternatives.
How it connects
Connects evolutionary search to parent selection, archived candidates, and exploit-or-explore sampling.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Self-Improving Agents. The section links point to the parts of the book it draws on.