Research map / Reinforcement learning / Policy gradient and actor-critic
An Alternative to Variance: Gini Deviation for Risk-averse Policy Gradient
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 changed | Where | Measured | Effect |
|---|---|---|---|
| Use Gini deviation instead of total-return variance as the risk penalty | Mean-GD policy gradient (MG) · Modified Guarded Maze, goal reward 40 | optimal risk-aversion rate | Helped |
| Increase the goal reward from 20 to 40 while keeping the risk-averse path's return variance unchanged | MVO · Modified Guarded Maze | optimal risk-aversion rate | Hurt |
| Increase the goal reward from 20 to 40 while keeping the risk-averse path's return variance unchanged | Tamar · Modified Guarded Maze | optimal risk-aversion rate | Hurt |
| Increase the goal reward from 20 to 40 while keeping the risk-averse path's return variance unchanged | MVP · Modified Guarded Maze | optimal risk-aversion rate | Hurt |
| Increase the goal reward from 20 to 40 | MVPI · Modified Guarded Maze | optimal risk-averse path discovery | Hurt |
| Use Gini deviation as the risk penalty | MG · LunarLander | left-landing rate | Helped |
| Add a variance-based risk penalty | MVO, Tamar, and MVP · LunarLander | reasonable policy learning | Hurt |
| Use Gini deviation as the risk penalty | MG · Mujoco: InvertedPendulum, HalfCheetah, and Swimmer | return | Helped |
| Use per-step reward variance as the risk proxy | MVPI · Distance-based risky HalfCheetah and Swimmer | forward progress | Hurt |
| Use Gini deviation instead of variance in the risk objective | Mean-GD policy gradient · Reward-scale manipulation in Modified Guarded Maze | risk-averse policy learning | Helped |
Closest papers by meaning
- Policy Evaluation for Variance in Average Reward Reinforcement Learning (ICML 2024 · similarity 0.90)
- Risk-Conditioned Reinforcement Learning: A Generalized Approach for Adapting to Varying Risk Measures (AAAI 2024 · similarity 0.87)
- Risk-sensitive control as inference with Rényi divergence (NeurIPS 2024 · similarity 0.86)
- Efficient Policy Evaluation with Safety Constraint for Reinforcement Learning (ICLR 2025 · similarity 0.86)
- Pessimism Meets Risk: Risk-Sensitive Offline Reinforcement Learning (ICML 2024 · similarity 0.86)
- Averaging $n$-step Returns Reduces Variance in Reinforcement Learning (ICML 2024 · similarity 0.86)
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.