Supervising strong learners by amplifying weak experts

TL;DR

The paper introduces Iterated Amplification, a training strategy that builds complex problem solutions from easier subproblems without external rewards, enabling learning of complex behaviors.

Contribution

It proposes Iterated Amplification as a novel method for training strong learners by amplifying weak experts through iterative subproblem solving, avoiding reliance on external reward signals.

Findings

Efficiently learns complex behaviors in algorithmic environments.

Demonstrates the effectiveness of amplification without external rewards.

Builds on and extends ideas from Expert Iteration.

Abstract

Many real world learning tasks involve complex or hard-to-specify objectives, and using an easier-to-specify proxy can lead to poor performance or misaligned behavior. One solution is to have humans provide a training signal by demonstrating or judging performance, but this approach fails if the task is too complicated for a human to directly evaluate. We propose Iterated Amplification, an alternative training strategy which progressively builds up a training signal for difficult problems by combining solutions to easier subproblems. Iterated Amplification is closely related to Expert Iteration (Anthony et al., 2017; Silver et al., 2017), except that it uses no external reward function. We present results in algorithmic environments, showing that Iterated Amplification can efficiently learn complex behaviors.