AMULED: Framework for Ethical Decision-Making in Reinforcement Learning

A research paper describing the AMULED framework for addressing moral uncertainty in reinforcement learning. The framework integrates multiple ethical theories using large language models to provide moral feedback, evaluated across two environments with 50-run replicates. The PDF document was authored by Rohit K. Dubey and last updated on 2026-05-04.

Use Cases

• Implementing a task-agnostic ethical layer for RL agents based on the described multi-perspective moral feedback.
• Studying belief aggregation methods like Belief Jensen–Shannon Divergence and Dempster–Shafer Theory for resolving conflicting ethical signals.
• Benchmarking ethical behavior improvements in simulated environments such as 'Finding Milk' and 'Driving and Rescuing'.
• Fine-tuning pre-trained RL models using LLM-generated feedback as a scalable alternative to human supervision.

Strengths

• The framework is evaluated with 50-run replicates, suggesting statistical robustness.
• Results include specific performance metrics, such as a 63.1% increase in desirable actions and a 5.1% increase in path length in the 'Finding Milk' environment.
• The document is openly licensed under CC-BY-4.0, permitting reuse and distribution.

Limitations

• The dataset is a 1.0 MB PDF document; the underlying experimental data, columns, and sample rows are unavailable.
• Row count is unknown, which may limit suitability assessment for direct ML training.
• Performance depends on LLM reasoning quality and can degrade in spatially complex settings, as noted in the discussion.

Provenance

Source

Rohit K. Dubey, published on figshare.

Collection Method

Research methodology involving simulation in two environments, multiple LLM backbones, and alternative belief aggregation methods.

Time Range

The study period is not specified.

Freshness

Last updated 2026-05-04 05:30:21; freshness should be verified.

Geography

Spatial coverage is not specified.