Research map / Reinforcement learning / Visual and representation RL
Beyond Rewards in RL for Cyber Defence
Elizabeth Bates, Chris Hicks, Vasilios Mavroudis
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use sparse positive-negative reward instead of complex dense negative | PPO in Yawning Titan · Yawning Titan (extended actions) | ScoreGT: 6.21 -> 2.00 | Improved |
| use sparse positive reward instead of dense negative | PPO in Yawning Titan · Yawning Titan (extended actions) | ScoreGT: 6.29 -> 2.69 | Improved |
| use sparse negative reward instead of dense negative | PPO in Yawning Titan · Yawning Titan (extended actions) | ScoreGT: 6.29 -> 10.29 | Hurt |
| use sparse positive reward instead of default CAGE reward | PPO in MiniCAGE · MiniCAGE | ScoreGT: 1.41 -> 1.29 | Improved |
| use SPN instead of SP at 20 nodes | PPO in Yawning Titan · Yawning Titan 20 nodes | ScoreGT: 1.87 -> 0.69 | Improved |
| use SP instead of SPN at 10 nodes | PPO in Yawning Titan · Yawning Titan 10 nodes | ScoreGT: 0.85 -> 0.63 | Mixed |
| use SPN instead of CDN under random agent order | PPO in Yawning Titan · Yawning Titan (random order) | ScoreGT: 11.79 -> 4.50 | Improved |
| use SP instead of CDN when blue acts before red | PPO in Yawning Titan · Yawning Titan (blue then red) | ScoreGT: 2.93 -> 0.27 | Improved |
| use dense rewards instead of sparse rewards | PPO in Yawning Titan · Yawning Titan (extended actions) | training reliability D_T: 0.11 -> 2.33 (e-3) | Hurt |
| randomise agent order instead of red then blue | PPO with SP reward · Yawning Titan (extended actions) | ScoreGT: 0.90 -> 6.91 | Hurt |
| use SP instead of default CAGE reward on operational server impacts | PPO in MiniCAGE · MiniCAGE | operational server impacts per episode: 0.03 -> 0.28 | Hurt |
| use SP instead of default CAGE reward on enterprise privilege escalations | PPO in MiniCAGE · MiniCAGE | enterprise escalations per episode: 21.80 -> 1.31 | Improved |
Closest papers by meaning
- Exploring the Efficacy of Multi-Agent Reinforcement Learning for Autonomous Cyber Defence: A CAGE Challenge 4 Perspective (AAAI 2025 · similarity 0.89)
- Optimal Attack and Defense for Reinforcement Learning (AAAI 2024 · similarity 0.87)
- Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.86)
- R*: Efficient Reward Design via Reward Structure Evolution and Parameter Alignment Optimization with Large Language Models (ICML 2025 · similarity 0.85)
- The Distributional Reward Critic Framework for Reinforcement Learning Under Perturbed Rewards (AAAI 2025 · similarity 0.85)
- Reward Machines for Deep RL in Noisy and Uncertain Environments (NeurIPS 2024 · similarity 0.85)
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.
