A 3D Characterization Framework for Intelligent Sequential Decision Making

TL;DR

This paper introduces a 3D characterization framework to analyze AI methods for sequential decision making. The framework evaluates methods based on their projection to Markov decision processes, degree of autonomy, and skill and computational cost. It reveals differences in design choices and performance characteristics among graph-based, reinforcement learning, and large language model-based approaches.

Why It Matters

The framework enables consistent comparison of AI methods, which is crucial for evaluating their effectiveness in real-world applications. It also highlights the trade-offs between different design choices, such as autonomy and computational cost. This can inform the development of more efficient and effective AI systems.

Key Findings

• LLM-based methods shift complexity from architecture to inference-time verification
• Neurosolver and FBRL have lower memory and runtime costs than LLM-based methods
• The 3D characterization framework can be applied to various AI methods and problems