Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 4·5 min read
Inference-Time Policy Alignment for Fair Reinforcement Learning
TL;DR
A new framework for inference-time policy alignment in reinforcement learning allows pretrained agents to adapt to fairness objectives without retraining. This approach enhances welfare-based fairness while maintaining core task performance.
✦ Why It Matters
Implement inference-time policy shaping to enhance fairness in your RL applications without retraining.
Key Takeaways
- Utilize the multiplicative policy shaping framework for fair RL.
- Adjust action probabilities without retraining your RL agent.
- Apply this method across various RL domains for improved fairness.
- Preserve core task performance while enhancing fairness metrics.
How It Works
The multiplicative policy shaping framework modifies the action probabilities of a pretrained RL agent based on welfare scores that reflect fairness objectives. This adjustment occurs at inference time, allowing the agent to align with new fairness criteria without changing its underlying policy parameters.