Agents

Youtube search... ...Google search

• Game Theory
• Reinforcement Learning (RL)
• Policy Gradient (PG)
• Multi-Task Learning (MTL)
• SMART - Multi-Task Deep Neural Networks (MT-DNN)
• Deep Distributed Q Network Partial Observability
• ChatGPT | OpenAI
  • Use ChatGPT here
  • Chrome summarizer extension: ReaderGPT

• Natural Language Generation (NLG)
• MineDojo ... framework built on the popular Minecraft game for embodied agent research
  • MINEDOJO: Building Open-Ended Embodied Agents with Internet-Scale Knowledge | L. Fan, G. Wang, Y. Jiang, A. Mandlekar, Y. Yang, H. Zhu, A. Tang, D. Huang, Y. Zhu, A. Anandkumar ... https://minedojo.org/videos/mineclip_policy_learning.mp4

Reinforcement Learning (RL) aims to make an agent (our “model”) learn through the interaction with an environment (this can be either virtual or real). RL was firstly developed to adhere to Markov Decision Process (MDP)es. In this ambit, an agent is placed in a stochastic stationary environment and tries to learn a policy through a reward/punishment mechanism. In this scenario, it is proved the agent will converge to a satisfactory policy. However, if multiple agents are placed in the same environment, this condition is no longer true. In fact, before the learning of the agent was only dependent on the interaction between the agent and the environment, now it is also dependent on the interaction between agents

Multi-Agents Reinforcement Learning (MARL)

• Multi-Agent Reinforcement Learning: A Selective Overview of Theories and Algorithms | K. Zhang, Z. Yang, and T. Başar

In a similar vein, multi-agent RL also addresses sequential decision-making problems, but with more than one agent involved. In particular, both the evolution of the system state and the reward received by each agent are influenced by the joint actions of all agents. More intriguingly, each agent has its own long-term reward to optimize, which now becomes a function of the policies of all other agents.

• Markov/Stochastic Games
  • Cooperative Setting
  • Competitive Setting
  • Mixed Setting
• Extensive-Form Games

Challenges

Despite a general model with broad applications, MARL suffers from several challenges in theoretical analysis, in addition to those that arise in single-agent RL.

1. Non-Unique Learning Goals - Unlike single-agent RL, where the goal of the agent is to maximize the long-term return efficiently, the learning goals of MARL can be vague at times. ...Indeed, the goals that need to be considered in the analysis of MARL algorithms can be multi-dimensional ... is undoubtedly a reasonable solution concept in game theory, under the assumption that the agents are all rational, and are capable of perfectly reasoning and infinite mutual modeling of agents. However, with bounded rationality, the agents may only be able to perform finite mutual modeling
2. Non-Stationarity - multiple agents usually learn concurrently, causing the environment faced by each individual agent to be non-stationary. In particular, the action taken by one agent affects the reward of other opponent agents, and the evolution of the state. As a result, the learning agent is required to account for how the other agents behave and adapt to the joint behavior accordingly. This invalidates the stationarity assumption for establishing the convergence of single-agent RL algorithms
3. Scalability Issue - To handle non-stationarity, each individual agent may need to account for the joint action space, whose dimension increases exponentially with the number of agents. This is also referred to as the combinatorial nature of MARL
4. Various Information Structures - Compared to the single-agent case, the information structure of MARL, namely, who knows what at the training and execution, is more involved. For example, in the framework of Markov games, it suffices to observe the instantaneous state st , in order for each agent to make decisions