Safe on Average, Unsafe in the Tail: When Is the Episodic-Cost Tail Controllable?

Key points

• In this work, we measure the episodic-cost tail using $CVaR{0.1}$, the average cost of the worst $10%$ of episodes.1
• We classify a policy as tail-safe when $CVaR{0.1}$ is within the safety budget.1
• To identify tail-unsafe policies, we evaluate five standard algorithms on three Safety-Gymnasium navigation tasks.1
• We then examine four constraint families on dense-hazard navigation and assess tail control across four navigation and four locomotion tasks.1

Sources (1)

• [1]Safe on Average, Unsafe in the Tail: When Is the Episodic-Cost Tail Controllable?
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 06:07 AM

  Safe reinforcement learning seeks policies that maximize return while satisfying constraints on cumulative cost.

  In this work, we measure the episodic-cost tail using $CVaR_{0.1}$, the average cost of the worst $10%$ of episodes.

Extractive summary: sentences quoted from the sources.