Work

A Conceptual Framework for Performance Evaluation and Classification of Multi-Agent Hybrid Game Problems Based on Reinforcement Learning: Impact Factors and Evaluation Metrics

Year: 2025

Type: article

Abstract: Multi-agent reinforcement learning (MARL) method for hybrid game is characterized by the coexistence of cooperativeness and competitiveness. This method has become more applicable for their complexity and wide applications in fields of architectural optimization, urban resource scheduling, and intelligent environmental control. However, the absence of a systematic framework for MARL algorithms’ assessment limits practical deployment in such fields. To address this concern, this paper intends to conduct a critical literature review to delve into the decisive factors of the scenarios and further develop conceptual framework for the subdivision classification, algorithm design and evaluation of hybrid games. The core research goal is to expound the differentiated classification system of hybrid game and the key determinants that significantly affect the performance of multi-agent reinforcement learning algorithms. Therefore, this paper proposes an evaluation conceptual framework for classification of hybrid games based on objectives and application scenarios, providing a reference for developers to select and design appropriate MARL algorithms and evaluate their performance according to different types of cooperative games.

Source: Urban and Building Science

DOI: 10.53941/ubs.2025.100004

Authors: Lihua Zhang, Xiaolei Yuan, Wenqi Jia, Angelica Elhokayem

Institutions: Tsinghua Shenzhen International Graduate School, Tsinghua University, Aalto University, The University of Texas at Arlington, Free University of Bozen-Bolzano

Language: English

Topic: Reinforcement Learning in Robotics

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: diamond