Toward Virtuous Reinforcement Learning

Summary

This seeded note is based on the first page and abstract. The paper critiques rule-based and scalar-reward ethical RL approaches and proposes a virtue-focused roadmap based on durable policy-level dispositions rather than isolated rule checks.

Key Claims

• Constraint-heavy deontological approaches can become brittle under ambiguity and non-stationarity.
• Scalarized reward approaches can hide moral trade-offs and invite proxy gaming.
• A virtue-oriented framing should evaluate stable traits, social learning, and explicit trade-off reporting.

Methods / Formalism

• Conceptual roadmap rather than a single algorithm.
• Emphasis on social learning, multi-objective and constrained RL, updateable priors, and explicit value assumptions.

Evidence / Experiments

• This appears to be a position and roadmap paper.
• Full support structure and related-work map still need full ingest.

Connections

• Useful counterpoint to strict Shielding-style safety methods.
• Connects ethics, Strategic Reasoning, and long-run adaptation themes near Continual Learning.
• Relevant to how the wiki should distinguish hard guarantees from normative modeling.

Open Questions

• How should virtue-style dispositions be represented in formal multi-agent settings?
• Which parts can be combined with formal constraints rather than replacing them?

Citation

Ghasemi, M., and Crowley, M. (2025). Toward Virtuous Reinforcement Learning: A Critique and Roadmap.