feat: simple mwem

[markusdemedeiros]

No description provided.

[markusdemedeiros]

Relevant test output:

[mwem] testing float-valued MWEM, (1 / 1)-DP per primitive, 5 rounds → 10 * 1/1 total
data: 300 records over 10 bins, biased toward bin 3
workload: 10 indicator queries (one per bin)

#0 transcript: [(3, 201), (3, 200), (3, 200), (3, 200), (3, 199)]
#1 transcript: [(3, 201), (3, 201), (3, 200), (3, 200), (3, 201)]
#2 transcript: [(3, 202), (3, 202), (3, 201), (3, 198), (3, 200)]

[mwem-hardt] small-scale 1D range query experiment in the spirit of Hardt 2012, Section 3.1
  domain: 15 bins
  workload: 12 1D range queries
  data: 200 synthetic records
  T = 4, ε₁/ε₂ = 1/1 per primitive

  trial 0: running mwem...
  trial 0: transcript = [(11, 146), (11, 147), (10, 145), (10, 146)]
  trial 0: replaying updater...
  trial 0: evaluating workload...
#0 avg squared error per query: 193.409108
  trial 1: running mwem...
  trial 1: transcript = [(10, 144), (11, 146), (10, 147), (10, 145)]
  trial 1: replaying updater...
  trial 1: evaluating workload...
#1 avg squared error per query: 200.200856
  trial 2: running mwem...
  trial 2: transcript = [(10, 145), (10, 143), (11, 144), (11, 149)]
  trial 2: replaying updater...
  trial 2: evaluating workload...
#2 avg squared error per query: 196.923975
mean over 3 trials: 196.844646
(units and parameter scale differ from Hardt's Figure 2; not a direct numerical comparison)