Principle:Ucbepic Docetl Pipeline Optimization Search

Knowledge Sources	MOAR DocETL Docs DocETL
Domains	Optimization, Search_Algorithms
Last Updated	2026-02-08 01:40 GMT

Overview

A search algorithm principle that uses Monte Carlo Tree Search (MCTS) to explore pipeline rewrite candidates, optimizing for accuracy and cost on a Pareto frontier.

Description

Pipeline Optimization Search applies MCTS to the problem of automatically rewriting LLM pipeline operations for better accuracy and lower cost. The MOAR (Multi-Objective Agent-based Rewriting) algorithm:

1. Selection: Choose a promising node in the search tree using UCB (Upper Confidence Bound)
2. Expansion: Apply a rewrite directive (from 25+ available) to generate a new pipeline variant
3. Simulation: Execute the variant pipeline on a sample dataset and evaluate accuracy
4. Backpropagation: Update node values throughout the tree based on the result

The search maintains a Pareto frontier tracking the tradeoff between accuracy and cost. Rewrite directives include operation chaining, gleaning, chunking, model swapping, fusion, compression, and more.

Usage

This principle is applied when running the MOAR optimizer (docetl build --optimizer moar). It requires an evaluation function to score pipeline quality and a set of available LLM models to search over.

Theoretical Basis

MCTS for pipeline optimization:

$UCB(n)={\overline{X}}_n+c\sqrt{\frac{\ln N}{n_i}}$

Where ${\overline{X}}_n$ is the average reward (accuracy change), $N$ is the parent visit count, $n_i$ is the node visit count, and $c$ is the exploration constant.

The Pareto frontier is maintained using hypervolume indicator calculations to track cost-accuracy tradeoffs across all explored pipeline variants.

Related Pages

Implemented By

• Implementation:Ucbepic_Docetl_MOARSearch_Search