Abstract
Current deep reinforcement learning (RL) algorithms are still highly task-specific and lack the ability to generalize to new environments. Lifelong learning (LLL), however, aims at solving multiple tasks sequentially by efficiently transferring and using knowledge between tasks. Despite a surge of interest in lifelong RL in recent years, the lack of a realistic testbed makes robust evaluation of LLL algorithms difficult. Multi-agent RL (MARL), on the other hand, can be seen as a natural scenario for lifelong RL due to its inherent non-stationarity, since the agents' policies change over time. In this work, we introduce a multi-agent lifelong learning testbed that supports both zero-shot and few-shot settings. Our setup is based on Hanabi -- a partially-observable, fully cooperative multi-agent game that has been shown to be challenging for zero-shot coordination. Its large strategy space makes it a desirable environment for lifelong RL tasks. We evaluate several recent MARL methods, and benchmark state-of-the-art LLL algorithms in limited memory and computation regimes to shed light on their strengths and weaknesses. This continual learning paradigm also provides us with a pragmatic way of going beyond centralized training which is the most commonly used training protocol in MARL. We empirically show that the agents trained in our setup are able to coordinate well with unseen agents, without any additional assumptions made by previous works. The code and all pre-trained models are available at https://github.com/chandar-lab/Lifelong-Hanabi.
Introduction
Lifelong Learning (LLL) aims to train models on a sequence of tasks without suffering from catastrophic forgetting. However, the LLL community currently lacks realistic benchmarks, often relying on synthetic modifications like Permuted MNIST or artificially tweaked physics in standard RL environments. Furthermore, quantifying the similarity between sequential tasks in these settings is often difficult.
Simultaneously, Multi-Agent RL (MARL) faces the challenge of "ad-hoc teamwork"—the ability to cooperate with novel partners zero-shot. Agents trained via standard Self-Play often develop brittle, idiosyncratic conventions that fail when paired with strangers. This research demonstrates that these two fields can solve each other's problems: MARL provides a naturally non-stationary environment for LLL, while LLL provides a framework to train MARL agents that generalize better to unseen partners.
The Lifelong Hanabi Framework
The authors propose a three-phase setup to evaluate LLL algorithms using the cooperative card game Hanabi:
- Pre-training: A massive pool of over 100 diverse agents is generated using various MARL algorithms (e.g., VDN, SAD, Other-Play) and architectures.
- Continual Training: A single "learner" agent is trained sequentially with partners sampled from the pool. Each new partner represents a new "task," requiring the learner to adapt while retaining past knowledge.
- Testing: The learner is evaluated on zero-shot and few-shot coordination with unseen agents to measure out-of-distribution generalization.
Quantifying Task Similarity with Cross-Play
A major advantage of this benchmark is that task similarity is easily quantifiable via a Cross-Play (CP) matrix. This grid shows the score of every pre-trained agent paired with every other agent. Low CP scores indicate agents are far apart in strategy space, allowing researchers to mathematically define "easy" and "hard" continual learning sequences based on partner diversity. This provides a rigorous way to evaluate how well LLL algorithms handle varying degrees of distribution shift.
Key Insights and Results
The benchmarking of standard LLL algorithms like Experience Replay (ER), A-GEM, and Elastic Weight Consolidation (EWC) yielded several critical insights:
- Zero-Shot Coordination Breakthrough: An Independent Q-Learning (IQL) agent trained continually with Experience Replay (IQL+ER) achieved a massive +8.09 improvement in coordination with completely different MARL algorithms. It outperformed specialized zero-shot coordination algorithms like SAD and Other-Play without needing their restrictive assumptions.
- The Optimizer Trade-off: The choice of optimizer is crucial in LLL. Adam allows for very fast adaptation to new tasks (ideal for few-shot learning) but suffers from severe catastrophic forgetting. SGD adapts slower but retains past knowledge significantly better.
- Continual Learning as a Regularizer: By forcing an agent to sequentially adapt to different partners, the agent learns robust, universally understandable conventions rather than brittle, specialized ones. This process acts similarly to Invariant Risk Minimization (IRM).
- Prioritized Replay: For memory-based LLL, prioritizing experiences based on TD-error (surprise) outperforms standard FIFO or reward-based sampling.
Limitations and Future Work
The current benchmark uses fixed pre-trained partners during the continual learning phase. Extending this to scenarios where partners are also learning simultaneously (co-adaptation) remains a significant challenge. Future work will also focus on human evaluation to see how well these agents coordinate with actual people and leveraging Meta-RL to enable even faster adaptation to new partners.
Summary updated