Meso Research Plan for Alpha Beta Filter

tags: Meso-Research

Meso Research Plans are living documents that indicates a general area of inquiry around one specific topic. The enumerated items are not exhaustive and will be added over time.

Context

• Filecoin's Block Reward as of now requires having estimators for what would be a sector 20-day block reward and also for what would be it's share of the Consensus Power.
  • Especifically, both the Network Block Reward and Network Quality-Adjusted Power are being estimated on each epoch.
  • The role of those estimators is to "smooth" the per epoch fluctuations on those values.
• Those estimators are implemented as being Alpha Beta filters, and they share the same parameter values even though they have different units, scale and underlying dynamics.
• It is known that Alpha Beta filters can have regions of stability, where parameters that are too low can lead to chaotic regimes, and parameters that are too high can lead to over-reaction.
• The performance and potential impacts of the filter were not studied exhaustively during implementation, and further theoretical / empirical evaluation can be helpful to understanding potential failure modes.

Why we care

• The alpha-beta filter introduces an additional complexity when representing the system on an aggregated formalism, as there's no trivial way of designing an equivalent representation a different time scale.
• Wrong values of the parameters can lead to unexpected network behaviour
  • This can include under-smoothing / over-reactions, therefore defeating the filter purpose while introducing potential attack vectors.
  • Another possibility is over-smoothing, introducing an inertia on the system that can introduce chaotic modes or under-performance of the block reward mechanism.
• Dynamics of the Alpha Beta filter and how to react to it can be non-trivial, especially on a governance setting.

Knowledge Applications

• Related to Filecoin as implementation
  • To determine success metrics for the filter implementation
• Related to Macrodynamics Digital Twin
  • Better backtesting performance especially on edge cases
  • Better knowledge about the extrapolation validity constraints of any representation for the filter
• Related to Governance
  • Better public awareness of the underlying assumptions and implications
  • Pre-emptive understanding of the consequence pathways when changing mechanisms related to Block Reward

Directions

• Empirical investigation through historical EDA and toy simulations of the stability/performance of the current parameters
• Determining best practices and constraints for translating the epoch-based filter into aggregated time estimators
• Determining performance metrics for the filter and their temporal distribution under a variety of past circumstances