Chapter 1: SELA and tree-search AutoML
Same as above in plain words: MCTS walks the tree using rollouts and validation scores to choose which branch to try next; UCT-DP changes the UCT score used when picking the next node so that deep, expensive training steps are less often pushed aside by shallow exploration.
What is Monte Carlo Tree Search (MCTS)?
In short: future experiments are arranged as a tree, and the same four steps repeat.
- ① Pick (selection): rules like UCT choose which node to visit next.
- ② Grow (expansion): attach a new child node (a new try) if it did not exist.
- ③ Roll (rollout): run code or a simulation on that branch to get a validation score.
- ④ Push up (backpropagation): send that score up to parents to update visit counts and averages.
SELA explores LLM-proposed pipeline branches with these four steps and validation scores.
What is UCT? (Upper Confidence Bound applied to trees) It is the scoring rule for choosing which sibling child to visit next. It mixes average reward so far (exploit) with how under‑visited a branch is (explore) in one formula, so you pick the next node by comparing numbers. The paper’s UCT-DP tweaks this UCT so deep training-heavy steps are not always behind shallow search.
Four steps (one cycle)
① Pick② Grow③ Roll④ Push up
Purple dashed line: one example path. Repeated runs accumulate scores on each branch.
[Abstract & intro] Three-line summary
Summary
- LLM limits: Code is often low-diversity and fails to converge to a good solution.
- Classical AutoML: Fixed pipelines (e.g., Auto-sklearn-style) resist dynamic reconfiguration when tasks change.
- SELA: Represent pipelines as a tree, schedule experiments with MCTS, use validation scores as feedback. UCT-DP biases search toward deep, training-heavy nodes.
Analogy: Following only the factory service playbook ≈ classical AutoML. Changing suspension, engine map, and tire pressure all at once and doing a single lap ≈ one-shot LLM codegen. SELA is like a race engineer who reads sector times and telemetry (validation scores) and branches on what to tune next.