# Fast and Slow Planning

**Authors:** Francesco Fabiano, Vishal Pallagani, Marianna Bergamaschi Ganapini, Lior Horesh, Andrea Loreggia, Keerthiram Murugesan, Francesca Rossi, Biplav Srivastava  
**arXiv:** [2303.04283](https://arxiv.org/abs/2303.04283)  
**Date:** 2023-03-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.04283v1/

## Key points

- Artificial Intelligence (AI) is increasingly prevalent in our daily lives
- AI-based systems lack natural intelligence capabilities
- Researchers developed an architecture inspired by D. Kahneman's cognitive theory called Thinking Fast and Slow
- The architecture is part of a broader AI paradigm called SOFAI (Slow and Fast AI)
- SOFAI utilizes multiple solving approaches with varying capabilities characterized as either fast or slow, including a metacognitive module to regulate these approaches according to human reasoning processes
- The team compared the new architecture with state-of-the-art solvers such as Fast Downward Helmert [2006] and EFP 2.0 Fabiano et al. [2020]
- Experiments were conducted on two domains: Blocks-World (BW) domain for classical planning and Coin in the Box (CB) domain for multi-agent epistemic planning
- The newly introduced system presented better results in terms of generality, solving a wider set of problems with an acceptable trade-off between solving times and solution accuracy compared to state-of-the-art solvers
- The team intends to explore ways to automatically tune the internal parameters of the architecture for optimal performance

## Summary

In recent years, Artificial Intelligence (AI) has become increasingly prevalent in various aspects of our daily lives. To address the issue of AI-based systems lacking the capabilities that we would naturally associate with intelligence, a team of researchers developed an architecture inspired by D. Kahneman's cognitive theory known as Thinking Fast and Slow. This architecture is part of a broader AI paradigm called SOFAI (Slow and Fast AI), which utilizes multiple solving approaches with varying capabilities characterized as either fast or slow. It also includes a metacognitive module to regulate these approaches in order to reflect human reasoning processes according to D. Kahneman's theory. To evaluate the effectiveness of their new architecture, the team compared it with state-of-the-art solvers such as Fast Downward Helmert [2006] and EFP 2.0 Fabiano et al. [2020]. They conducted experiments on two distinct domains: Blocks-World (BW) domain for classical planning and Coin in the Box (CB) domain for multi-agent epistemic planning. The results showed that the newly introduced system presented better results in terms of generality, solving a wider set of problems with an acceptable trade-off between solving times and solution accuracy compared to state-of-the-art solvers. However, the team intends to explore ways to automatically tune the internal parameters of the architecture for optimal performance. In conclusion, this work presents a promising approach to enhancing AI-based planning systems' capabilities by incorporating cognitive theory and utilizing multiple solving approaches with a metacognitive module. The ongoing work aims to further refine and optimize this architecture for even better performance.