ZAC-Training: Collective Provisioning of Model Training and Updates in Heterogeneous Agent Populations

@misc{zhang2026zactraining,
  title = {{ZAC-Training: Collective Provisioning of Model Training and Updates in Heterogeneous Agent Populations}},
  author = {Jincheng Zhang},
  howpublished = {Zenodo},
  year = {2026},
  month = oct,
  doi = {10.5281/zenodo.23450968},
  url = {https://zenodo.org/records/23450968},
  note = {CC BY 4.0},
}

TY  - UNPB
TI  - ZAC-Training: Collective Provisioning of Model Training and Updates in Heterogeneous Agent Populations
AU  - Jincheng Zhang
PY  - 2026
DA  - 2026/10/18
PB  - Zenodo
DO  - 10.5281/zenodo.23450968
UR  - https://zenodo.org/records/23450968
AB  - Machine learning systems deployed across heterogeneous agent populations are commonly designed around the assumption that individual agents must possess sufficient computational resources to train, fine-tune, or update their own models, or that training must be performed by a dedicated centralized infrastructure. However, many agents primarily generate data, execute inference, perform physical tasks, and provide operational feedback, while only a subset may have the computational capacity, energy budget, or infrastructure required for model training. Requiring every agent to provision independent training capabilities may therefore create unnecessary hardware requirements, duplicated infrastructure, and underutilized computational resources.  This paper proposes ZAC-Training, a research framework for collective provisioning of model training and model-update services within heterogeneous agent populations. Under this framework, training capability is treated as a collectively allocatable system resource rather than a mandatory local capability of every participating agent. Selected nodes may perform model training, fine-tuning, adaptation, validation, or update generation, while other nodes contribute data, observations, feedback, evaluation signals, or operational requirements without independently executing the complete training process. Validated model artifacts or compatible updates can subsequently be distributed to eligible recipients.  The framework distinguishes training-service provisioning from data contribution, update aggregation, model distribution, and local adaptation. It formulates a joint optimization problem involving training-node selection, workload assignment, communication cost, model quality, data heterogeneity, update freshness, reliability, privacy, and security. It also introduces a provisioning-coverage measure to characterize the proportion of participating agents that can obtain required model-update services without independently provisioning the designated training capability. A conditional break-even analysis illustrates when collective training may be preferable to independent training infrastructure.  ZAC-Training is positioned in relation to federated learning, distributed training, edge learning, and collaborative model adaptation. These established areas already address many components of collective training. The proposed research contribution is therefore the explicit treatment of training-capability provisioning as a joint system-design decision, including which agents need local training resources, which can rely on external training services, and how the resulting dependencies affect total system cost and operational feasibility. The framework is conceptual and requires comparative experiments to establish its practical value.
ER  -