Scholarlink

In multi-teacher reinforcement learning, conventional wisdom suggests that combining expert knowledge through ensemble methods should improve performance. We reveal a striking paradox: in environments with changing goals, ensemble methods that achieve the highest agreement among teachers deliver the worst performance (32.3% success rate) – even worse than random teacher selection (34.5%). Through controlled experiments in a drifting grid world where four expert teachers guide a learning agent, we demonstrate that confidence-weighted voting creates false consensus by amplifying outdated expertise. Our analysis of 30 random seeds (F = 8957.6, p < 0.0001) shows that when environments change, teacher disagreement is not noise to be reduced but a valuable signal of adaptation. We introduce the Teacher Confusion Index (TCI) and Goal Coherence Score (GCS) to quantify this phenomenon, revealing a positive correlation (r = 0.277) between disagreement and performance. These findings challenge fundamental assumptions about ensemble learning in non-stationary environments, with implications for any multi-expert system facing concept drift.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The Consensus Paradox: When Low Disagreement Leads to Catastrophic Failure in Multi-teacher Reinforcement Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Maher Mesto,
  • Francisco Cruz

摘要

In multi-teacher reinforcement learning, conventional wisdom suggests that combining expert knowledge through ensemble methods should improve performance. We reveal a striking paradox: in environments with changing goals, ensemble methods that achieve the highest agreement among teachers deliver the worst performance (32.3% success rate) – even worse than random teacher selection (34.5%). Through controlled experiments in a drifting grid world where four expert teachers guide a learning agent, we demonstrate that confidence-weighted voting creates false consensus by amplifying outdated expertise. Our analysis of 30 random seeds (F = 8957.6, p < 0.0001) shows that when environments change, teacher disagreement is not noise to be reduced but a valuable signal of adaptation. We introduce the Teacher Confusion Index (TCI) and Goal Coherence Score (GCS) to quantify this phenomenon, revealing a positive correlation (r = 0.277) between disagreement and performance. These findings challenge fundamental assumptions about ensemble learning in non-stationary environments, with implications for any multi-expert system facing concept drift.