# Meta-operators for Enabling Parallel Planning Using Deep Reinforcement Learning

**Authors:** Ángel Aso-Mollar, Eva Onaindia  
**arXiv:** [2403.08910](https://arxiv.org/abs/2403.08910)  
**Date:** 2024-03-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.08910v1/

## Key points

- Growing interest in leveraging Reinforcement Learning (RL) techniques for developing general policies in AI planning
- Introduction of innovative concept of meta-operators and exploration of their potential for enhancing planning outcomes
- Incorporating meta-operators into RL action space allows for new perspectives like parallel planning to be explored
- Analysis of performance and complexity implications of including meta-operators in the RL process
- Focus on redefining RL action space to align more closely with the planning perspective in domains where traditional generalized planning models have not yielded satisfactory outcomes
- Delving into fundamental concepts such as Planning and RL integration, including Generalized Planning
- Reinforcement Learning involves learning optimal decisions through trial and error interactions with an environment
- Inclusion of meta-operators in RL enables parallel planning and reduces plan length compared to traditional approaches
- Conducting various experiments to demonstrate the effectiveness of meta-operators in improving planning outcomes
- Exploration of whether adopting this approach offers benefits over other models lacking meta-operator integration

## Summary

In the realm of Artificial Intelligence (AI) planning, there is a growing interest in leveraging Reinforcement Learning (RL) techniques to develop general policies. This paper introduces the innovative concept of meta-operators and explores their potential for enhancing planning outcomes. Traditionally, AI planning transition models have been mapped to Markov Decision Processes using a one-to-one correspondence of action spaces. However, incorporating meta-operators into the RL action space allows for new perspectives such as parallel planning to be explored. The primary objective of this research is to analyze the performance and complexity implications of including meta-operators in the RL process. It focuses on domains where conventional generalized planning models have not yielded satisfactory outcomes and aims to redefine the RL action space in a way that aligns more closely with the planning perspective. The paper delves into fundamental concepts such as Planning and RL integration, including Generalized Planning. Classical planning is defined as finding a sequence of actions that lead from an initial state to achieving specific goals. A planning domain consists of fluents describing properties and action schemas defining actions' preconditions, effects, and negative effects. Reinforcement Learning involves learning optimal decisions through trial and error interactions with an environment. The inclusion of meta-operators in RL enables parallel planning and reduces plan length compared to traditional approaches. The structure of this paper includes defining key concepts like meta-operators and integrating them into the RL framework. Various experiments are conducted to demonstrate the effectiveness of meta-operators in improving planning outcomes. The discussion explores whether adopting this approach offers benefits over other models lacking meta-operator integration. Overall, this study aims to pave the way for redefining RL action spaces by introducing meta-operators for enhanced planning capabilities. Through empirical analysis and theoretical discussions, it seeks to advance AI planning methodologies towards more efficient and effective decision-making processes in complex environments.