Abstract:This paper focuses on the dilemma faced by largescale multi-agent systems. With the increase of agent size, the policy search space grows exponentially, reaching the task complexity that is difficult for the current mainstream multiagent algorithms to deal with, so that the optimal strategy is converged in large-scale multi-agent systems. In order to make the learning process of large-scale multi-agent tasks more smooth, many researchers have introduced the method of course learning, setting the scale from small to large courses, compared with learning training from scratch, this way can continue to inherit the knowledge learned in the smaller scale tasks, making the training smoother. However, in order to realize the multiagent course learning from small to large scale, we face the following difficulties. First, the neural network structure we designed can adapt to the observation input of agents of different scales, rather than only applicable to a fixed scale. Second, in the multi-stage course learning, due to the network's inefficient representation of multi-agent systems, the knowledge learned in small-scale courses cannot be effectively utilized in large-scale courses, and even the problem of catastrophic forgetting occurs in the neural network. Specifically, we view multi-agent systems from a graph perspective and derive the Graph Decomposition Invariant (GDI) for the first time. We prove that neural network architectures satisfying the GDI are of great help in realizing efficient course learning. We design a model architecture called GDI model suitable for multi-agent curriculum learning, which satisfies the graph decomposition invariant proposed by us. Finally, we designed small-scale to large-scale multi-agent tasks on two experimental platforms, StarCraft and Neural MMO, and conducted a large number of comparative experiments, which verified that our proposed GDI model has a great promotion effect on course learning.

Developing Parallel Dependency Graph In Improving Game Balancing

Dynamic Resource Access Using Graphical Game in Asymmetric Wireless Networks.

Metagame Autobalancing for Competitive Multiplayer Games

A Game of Attribute Decomposition for Software Architecture Design

Load Balancing for Distributed Intelligent Edge Computing: A State-Based Game Approach.

Hybrid fighting game AI using a genetic algorithm and Monte Carlo tree search

Design and implementation of a survival game towards achieving user's satisfaction by providing balanced difficulty

Player Dominance Adjustment in Games

Exploring Equilibrium Strategies in Network Games with Generative AI

Research on Dynamic Graph Layout by Parallel Computing and Markov Process.

Identifying and Clustering Counter Relationships of Team Compositions in PvP Games for Efficient Balance Analysis

Navigating the Landscape of Multiplayer Games

Intelligent Generation of Graphical Game Assets: A Conceptual Framework and Systematic Review of the State of the Art

Challenges and Solutions of the Application of Data-Driven Technology in Network 3D Game Design

Scaling Up Multi-Agent Reinforcement Learning Via Graph Decomposition Invariant Network

Dependency Structure Search Bayesian Optimization for Decision Making Models

Optimizing Software Development Requirements Based on Dependency Relations.

An approach to analyzing dependency of concurrent programs

Automatically Reinforcing a Game AI

GEEvo: Game Economy Generation and Balancing with Evolutionary Algorithms

A Survey on Distributed Online Optimization and Game