This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
To address the generalization problem of reinforcement learning (RL) agents, this paper proposes a Dual-System Adaptive Decision Framework (DSADF), consisting of System 1 (an RL agent and memory-based intuitive decision-making) and System 2 (a Vision Language Model (VLM)-based deep reasoning). Inspired by Kahneman's Systems 1 and 2 theories, it combines the rapid response capabilities of RL agents with the inference capabilities of VLMs to enable efficient and adaptive decision-making in complex environments. Experimental results in video game environments (Crafter and Housekeep) demonstrate that DSADF outperforms existing methods on both unknown and known tasks.
Takeaways, Limitations
•
Takeaways:
◦
A novel approach to solving the generalization problem of RL agents is presented.
◦
Demonstrating the potential for efficient and adaptive decision-making through the combination of System 1 and System 2.
◦
Improving the Performance of RL Agents Using VLM
◦
Presenting the possibility of competent decision-making in complex environments
•
Limitations:
◦
The performance of the proposed framework may be limited to specific gaming environments.
◦
Further research is needed on the interactions and coordination mechanisms between System 1 and System 2.
◦
Verification of generalization performance in more diverse and complex environments is needed.
◦
Consideration of VLM's computational cost and training data dependency issues is necessary.
