Dynamic Encryption Windows backtest

[rad2520]

goal:
backtest the DEW algorithm on real historic subnet metagraph data to observe behavior for different subnet types under different parameter settings.

open/close encryption window is just opening/closing measurement windows of the backtest and resetting Dc(e) and Dd(e).

pick starting epoch, then run over all epochs up to the latest. starting epoch should be as early in the subnets history as possible, where reasonably large validator and miner population has joined the subnet and consensus stabilized.

run on the major Bittensor subnets, due to significantly higher diversity and activity volume i.e.

• s1, s8, s9, s18, s19, s24, s25

run on functional Commune subnets with sufficient history i.e. Synthia, OpenScope, Mosaic

creating a rich dataset of real encryption windows to derive stats and analysis

initial parameters:

• p = 0.05
• max_encryption_duration = off

analysis ideas

• plot average encryption window length with p = 0.025, 0.03, 0.035, 0.04 etc up to 0.15
• plot x axis encryption windows, y axis encryption length to visualize outliers and to see if there are any noticeable patterns. this can be used for picking the max_encrpytion_duration parameter individually per subnet
• calculate gini coefficient for distribution of portions of the final percent difference (approx p) between Dc(e) and Dd(e) across epochs contained per encryption window and plot it x axis encryption windows y axis gini coefficient
  -> this will give us an easy to interpret view of how evenly copier reward penalty is distributed across epochs e.g. if its usually the case that 1-2 epochs make up 70% of the reward penalty in a 10 epoch encryption window
• get average encryption length for every relevant-enough BT subnet with p=0 and 9% delegation fee