Advancements in Reinforcement Learning and Multi-Agent Systems

The field of reinforcement learning and multi-agent systems is rapidly advancing, with a focus on developing more efficient, scalable, and interpretable methods. Recent developments have seen a shift towards hierarchical reinforcement learning, dynamic planning, and the use of large language models to improve agent performance. Notably, researchers are exploring the use of language-driven hierarchical task structures as explicit world models for multi-agent learning, which has shown promise in improving sample efficiency and enabling more sophisticated strategic behaviors. Noteworthy papers include: Scalable Option Learning, which proposes a highly scalable hierarchical RL algorithm that achieves a 25x higher throughput compared to existing hierarchical methods. OmniActor, a generalist agent that leverages the synergy between GUI and embodied data to outperform agents trained on single modalities. PolicyEvolve, a framework for generating programmatic policies in multi-player games, which reduces reliance on manually crafted policy code and achieves high-performance policies with minimal environmental interactions.

Sources

Scalable Option Learning in High-Throughput Environments

Dynamic Speculative Agent Planning

OmniActor: A Generalist GUI and Embodied Agent for 2D&3D Worlds

Learning General Policies From Examples

Design and Optimization of Reinforcement Learning-Based Agents in Text-Based Games

Learning When to Plan: Efficiently Allocating Test-Time Compute for LLM Agents

A Comprehensive Review of Multi-Agent Reinforcement Learning in Video Games

ProST: Progressive Sub-task Training for Pareto-Optimal Multi-agent Systems Using Small Language Models

Bootstrapping Task Spaces for Self-Improvement

Language-Driven Hierarchical Task Structures as Explicit World Models for Multi-Agent Learning

PolicyEvolve: Evolving Programmatic Policies by LLMs for multi-player games via Population-Based Training

PillagerBench: Benchmarking LLM-Based Agents in Competitive Minecraft Team Environments

AgentGym-RL: Training LLM Agents for Long-Horizon Decision Making through Multi-Turn Reinforcement Learning

Built with on top of