AI Paper Atlas

Research map / Reinforcement learning / MDPs and planning

Reinforcement Learning Can Be More Efficient with Multiple Rewards

Christoph Dann, Yishay Mansour, Mehryar Mohri

ICML 2023 · Reinforcement learning › MDPs and planning · approach: POMDPs and confounded RL

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add multiple reward functions (max operator, action elimination) instead of a single rewardmulti-armed bandit action elimination · Multi-armed bandits, theory (Theorem 2)regret bound per action scalingImproved
Add multiple reward functions in tabular MDP action eliminationmulti-reward AMB-style action elimination · Tabular MDPs, theory (Theorem 3 vs single-reward Xu et al. 2021 bound)regret bound scaling per state-action pairImproved
Use multiple rewards with n=m reward functions in the MDP example vs single reward (ablation)multi-reward RL on constructed MDP · Simulated MDP with H=n (Figure 2)cumulative regret on reward r1Improved
Add second reward function in two-reward bandit example (Example 3) (ablation)multi-armed bandit action elimination · Bandit Example 3 (theory)regret bound: decreases regret bound by a ln(T) factorImproved
Use multi-reward operator choice (mean/CVaR) instead of max when rewards disagreemulti-armed bandit action elimination · Theory (Theorem 1)operator regretMixed
Model selection over the operator (max_i or CVaR family) adaptivelymulti-armed bandit action elimination · Two-instance lower bound (theory)regretNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy