Skip to content
Skydvn-PNU
DG
FDG
Federated Segmentation
Federated Unlearning - Adversarial Recovery
Decentralized MARL
VLM
LLM-ComSoc
Diffusion Models
Protein Foldings
Reasoning/Symbolic: A Survey
Decentralized MARL: A survey
Resource Allocation - SemCom: A survey
Untitled page
SubGraph-FL
WorldModel-DRL
Share
Explore

icon picker
Decentralized MARL




⁠

Decentralized MARL
Paper Name
Categories
Type
Concept
Recent issues
Motivations
Contributions
Evaluation List
Target
Conf./Jour.
Year
Link
Coach-Player Multi-Agent Reinforcement Learning for Dynamic Team Composition
Centralized Training Decentralized Execution (CTDE).
Coach with full-view // players have partial-views
Coach can distribute information → agents with limited amount.
Most deep CTDE for cooperative MARL limited to a fixed number of homogeneous agents (C1).
Computationally prohibitive to re-train the agents (C2).
Agent can only access to its own decisions and partial environmental observations at test-time (C3).
All agents to communicate is too expensive in many scenarios (C4).
Generalize zero-shot to new compositions.
C3: Introducing communication
C4: Centralized coach → periodically distributes strategic information (full view).
Communication through continuous strategy vector.
Variational objective → regularize learning of the strategy.
Adaptive policy → coach only communicate if needed.
Strategy vector is encoded using VAE-based Encoder.
Resource Collection
Rescue Game
Starcraft Multi-agent Challenge (SMAC).
​
De-confounded Value Decomposition for Multi-Agent Reinforcement Learning
Centralized Training Decentralized Execution (CTDE).
- Credit Assignment: deduce contributions of individual agents from overall success.
​
​
​
​
On Improving Model-Free Algorithms for Decentralized Multi-Agent Reinforcement Learning
Decentralized Training.
Sample-efficient Model-free Algorithm.

Exponential dependence as it usually needs to exhaustively search the joint action space.
Computation bottleneck can be solved by communications → distributing workload.
Communication-based → communication overheads.
Stage-based V-Learning for General-Sum Markov Games.
Learning CCE
Learning CE
Learning NE in Markov Potential Games.
​
​
There are no rows in this table
⁠

Want to print your doc?
This is not the way.
Try clicking the ⋯ next to your doc name or using a keyboard shortcut (
CtrlP
) instead.