Fig 1.
High-level illustration of the workflow of our proposed mechanism.
During initial deployment, if no pretrained model is available, alternative methods assign roles to nodes. Later, supervisors apply ML models using different features to select the next supervisors and a group of block producers. Blocks are produced in a randomized order and the block production outcome is used to improve the ML model.
Fig 2.
High-level illustration of how PoPI uses its ML model to rank nodes and select a top-n group from which a randomized method determines their block production order.
Fig 3.
Illustration of the duties of supervisor nodes during their term.
Supervisors collect node features, run inference after block production, and broadcast results before handing off to the next set of supervisors.
Table 1.
Extended list of potential model features organized by category for PoPI.
Fig 4.
Evolution of consensus and their capabilities in block producer selection.
Table 2.
Comparative analysis of PoPI with existing ML-based consensus mechanisms.
Fig 5.
The block latency of consensus mechanisms with the increase of nodes.
Table 3.
Statistical summary of block latency at 100 runs for 1000 nodes.
Fig 6.
The block latency of consensus mechanisms in 100 runs for 1000 nodes.
Fig 7.
The throughput of consensus mechanisms with the increase of nodes.
Fig 8.
The computational overhead of consensus mechanisms with the increase of nodes.
Fig 9.
The active participation time of consensus mechanisms with the increase of nodes.
Table 4.
Ablation study of different components of PoPI for 1000 nodes.
Fig 10.
Analysis of fair participation in PoPI for 1000 nodes, showing (a) node participation percentage and (b) participation frequency.
Table 5.
Resource consumption of different supervised learning models.