Scholarlink

Sequential decision-making (SDM) is the process of choosing the best actions in a Gaming Environment, step by step over time. But it has many challenges, including that rewards are rare or misleading and Environments are different for each agent. Reinforcement Learning (RL) and Evolution Strategies (ES) are two popular ways to solve SDM problems. Since RL uses gradients and value-based learning and ES works like a black-box and uses a group of solutions (population-based), this paper compares RL and ES in both centralized and federated learning situations. It also uses simple games like Snake Game, Mountain Car, CartPole, and Assault (an Atari game) for testing. The study compares (1) how fast RL and ES learn (convergence) and (2) how much communication they need. This study helps us choose the right method for real-life cases with limited resources, privacy concerns, and decentralized learning environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis of Federated Learning (FL), Reinforcement Learning (RL) and Evolution Strategy (ES) in Gaming Environments

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Priyanshi Soni,
  • Tanmay Somani,
  • Shaligram Prajapat,
  • Harshil Soni

摘要

Sequential decision-making (SDM) is the process of choosing the best actions in a Gaming Environment, step by step over time. But it has many challenges, including that rewards are rare or misleading and Environments are different for each agent. Reinforcement Learning (RL) and Evolution Strategies (ES) are two popular ways to solve SDM problems. Since RL uses gradients and value-based learning and ES works like a black-box and uses a group of solutions (population-based), this paper compares RL and ES in both centralized and federated learning situations. It also uses simple games like Snake Game, Mountain Car, CartPole, and Assault (an Atari game) for testing. The study compares (1) how fast RL and ES learn (convergence) and (2) how much communication they need. This study helps us choose the right method for real-life cases with limited resources, privacy concerns, and decentralized learning environments.