Continual Learning
Overview
This area covers learning systems that update over time while preserving useful prior knowledge and adapting safely to change. Its current wiki role is a focused routing hub for long-lived adaptation, non-stationary multi-agent training, and explicit task-model revision; broad neural-network training dynamics remain in Deep Learning Fundamentals unless persistence over time is central.
Active Questions
- How should safety guarantees evolve when the policy itself keeps changing?
- Which adaptation questions are specific enough for this area, and which are better left to Deep Learning Fundamentals or Safe Multi-Agent Reinforcement Learning?
- What representations support stable transfer without catastrophic forgetting?
- How should long-lived agents update values, norms, or constraints under non-stationarity?
- How should an agent adapt when other agents’ policies, population size, or capabilities change during learning?
- How can an agent revise an explicit task model while preserving useful policy estimates?
- How should reusable temporal subtasks and policies carry forward into harder tasks?
Key Concepts
- Safe Reinforcement Learning
- Strategic Reasoning
- Multi-Agent Non-Stationarity
- Non-Markovian Reinforcement Learning
- Automata Learning
- Reward Machines
Key Sources
- Ghasemi2025 - Toward Virtuous Reinforcement Learning
- Zhang2025 - Trustworthy Reinforcement Learning under Constraints and Perturbations
- Alinejad2026 - Dynamic Automaton Refinement and Planning for Non-Markovian RL
- Furelos-Blanco2023 - Hierarchies of Reward Machines
- Papoudakis2019 - Dealing with Non-Stationarity in Multi-Agent Deep Reinforcement Learning