Succinct and Robust Multi-Agent Communication With Temporal Message Control
Authors: Sai Qian Zhang, Qi Zhang, Jieyu Lin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that TMC can significantly reduce inter-agent communication overhead without impacting accuracy. Furthermore, TMC demonstrates much better robustness against transmission loss than existing approaches in lossy networking environments. We have evaluated TMC on multiple environments including Star Craft Multi-Agent Challenge (SMAC) [25], Predator-Prey [19] and Cooperative Navigation [19]. |
| Researcher Affiliation | Collaboration | Sai Qian Zhang Harvard University Jieyu Lin University of Toronto Qi Zhang Microsoft |
| Pseudocode | Yes | Algorithm 1: Communication protocol at agent n |
| Open Source Code | Yes | A video demo is available at [3] to show TMC performance, and the code for this work is provided in [2]. and reference "[2] Tmc code repository. https://github.com/saizhang0218/TMC." |
| Open Datasets | Yes | We have evaluated TMC on multiple environments including Star Craft Multi-Agent Challenge (SMAC) [25], Predator-Prey [19] and Cooperative Navigation [19]. |
| Dataset Splits | No | All the algorithms are trained with ten million episodes. We pause the training process and save the model once every 500 training episodes, then run 20 test episodes to measure their winning rates. (This describes training and testing during training, but does not specify a distinct *validation* dataset split for hyperparameter tuning or early stopping criteria.) |
| Hardware Specification | No | Specifically, we use an IEEE 802.11ac Access Point (AP) and a pair of Raspberry Pi 3 models [1] as intermediate router, sender and receiver, respectively. (This describes hardware for simulating network conditions, not for training the models. No specific hardware for model training/inference is mentioned.) |
| Software Dependencies | No | The paper mentions software components like 'recurrent neural network', 'Gated Recurrent Unit (GRU)', 'Multilayer Perceptron (MLP)', but does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | For TMC hyperparameters, we apply ws = 6, β1 = 1.5, β2 = 1, β3 = 0.5 for all 6 scenarios. Other hyperparameters such as λr, λs used in equation 3, δ in Algorithm 1 are shown in the legends of Figure 4. For TMC, we apply ws = 4, δ = 0.02, λr = 0.3, β1 = 1.5, β2 = 1 and λs = 1.0. |