Continual Knowledge Adaptation for Reinforcement Learning

Jinwu Hu, Zihao Lian, Zhiqun Wen, Chenghao Li, Guohao Chen, Xutao Wen, Bin Xiao, Mingkui Tan
_{South China University of Technology, Pazhou Laboratory}

TL;DR: CKA-RL introduces a Continual Knowledge Adaptation framework for reinforcement learning that maintains a pool of task-specific knowledge vectors and dynamically reuses historical knowledge to adapt to new tasks.

---

✨ Abstract

Reinforcement Learning enables agents to learn optimal behaviors through interactions with environments. However, real-world environments are typically non-stationary, requiring agents to continuously adapt to new tasks and changing conditions. Although Continual Reinforcement Learning facilitates learning across multiple tasks, existing methods often suffer from catastrophic forgetting and inefficient knowledge utilization. To address these challenges, we propose Continual Knowledge Adaptation for Reinforcement Learning (CKA-RL), which enables the accumulation and effective utilization of historical knowledge. Specifically, we introduce a Continual Knowledge Adaptation strategy, which involves maintaining a task-specific knowledge vector pool and dynamically using historical knowledge to adapt the agent to new tasks. This process mitigates catastrophic forgetting and enables efficient knowledge transfer across tasks by preserving and adapting critical model parameters. Additionally, we propose an Adaptive Knowledge Merging mechanism that combines similar knowledge vectors to address scalability challenges, reducing memory requirements while ensuring the retention of essential knowledge. Experiments on three benchmarks demonstrate that CKA-RL outperforms state-of-the-art methods, achieving an improvement of 4.20% in overall performance and 8.02% in forward transfer.

---

🚀 Quick Start

1. Clone the repo

   git clone https://github.com/Fhujinwu/CKA-RL.git
   cd CKA-RL
   

2. Create environments

   💡 Install the environment only for the experiments you plan to run.
   The project provides two separate requirement files for different experiment groups:

   • experiments/atari/requirements.txt -- dependencies for Atari experiments
   • experiments/meta-world/requirements.txt –- dependencies for Meta-World experiments

   # === Environment setup for 🕹️ Atari experiments ===
   conda create -n cka-rl-atari python=3.10 -y
   conda activate cka-rl-atari
   pip install -r experiments/atari/requirements.txt
   
   # === Environment setup for 🌍 Meta-World experiments ===
   conda create -n cka-rl-meta python=3.10 -y
   conda activate cka-rl-meta
   pip install -r experiments/meta-world/requirements.txt
   

▶️ Usage

🕹️ Atari Experiments

1. Run CKA-RL on a specific Atari environment

   # Example: Run CKA-RL on the Atari environment "Freeway"
   ATARI_ENV="ALE/Freeway-v5"     # Options: ALE/Freeway-v5, ALE/SpaceInvaders-v5
   SEED=42
   
   cd experiments/atari/
   python run_experiments.py \
       --method_type CKA-RL \
       --env $ATARI_ENV \
       --first-mode 0 \
       --last-mode 7 \             # Use --last-mode 9 for SpaceInvaders
       --seed $SEED \
       --tag main
   

   🗂️ Training logs and results will be automatically saved to: ./experiments/atari/data/Freeway/main

2. Process Results

   ⚠️ Note: Before running the following scripts, ensure that all methods (including baselines) have been executed; otherwise, process_results.py may fail due to missing result files.

   # Step 1. Gather all results from experiment logs
   python gather_rt_results.py \
       --base_dir ./experiments/atari/data/Freeway/main \
       --env Freeway             # Options: Freeway, SpaceInvaders
   
   # Step 2. Process the aggregated results
   python process_results.py \
       --data-dir ./experiments/atari/data/Freeway/main
   

🌍 Meta-World Experiments

1. Run CKA-RL on Meta-World environment

   # Example: Run CKA-RL
   python run_experiments.py \
           --algorithm cka-rl \
           --tag main
   
   

   🗂️ Training logs and results will be automatically saved to: ./experiments/meta-world/runs/main

2. Process Results

   ⚠️ Note: Before running the following scripts, ensure that all methods (including baselines) have been executed; otherwise, process_results.py may fail due to missing result files.

   # Step 1. Gather all results from experiment logs
   python extract_results.py
   
   # Step 2. Process the aggregated results
   python process_results.py
   

💬 Citation

If you find our work useful, please consider citing:

@inproceedings{hu2025CKARL,
  title={Continual Knowledge Adaptation for Reinforcement Learning},
  author={Hu, Jinwu and Lian, Zihao and Wen, Zhiquan and Li, Chenghao and Chen, Guohao and Wen, Xutao and Xiao, Bin and Tan, Mingkui},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}

🙏 Acknowledgements

We gratefully acknowledge the following open-source contributions:

• componet for baseline codebase
• loss-of-plasticity for CbpNet implementation
• mask-lrl for MaskNet implementation
• crelu-pytorch for CReLU implementation

Our work builds upon these open-source efforts; we thank the authors for their valuable contributions.

Star History