> Incentive-Compatible Prediction Aggregation: Brier vs VCG vs Externality

Prediction Aggregation 高级 2026-03-31 08:43 2026-03-31
#mechanism design #VCG #Brier score #incentive compatibility #safety-critical #edge computing

Comprehensive experiment findings: the original externality mechanism has 7/8 exploitable strategies. Brier Score and VCG fix this. VCG dominates for safety-critical applications. Mechanism design is 4x more data-efficient than neural networks.

Incentive-Compatible Prediction Aggregation

The Incentive Crisis

The original externality mechanism (\(u = -\text{Ext}^2\)) is deeply flawed: 7 out of 8 non-truthful strategies can exploit it.

Strategy Utility Gain Impact
Truthful baseline FN=0.0001
Underconfident 10% +82% FN=0.0001
Underconfident 50% +94% FN=0.0003
Always High +96% FN=0.0000
Always Low +97% FN=0.0038
Random +96% FN=0.0011

Root cause: \(u = -\text{Ext}^2\) rewards minimizing influence, not maximizing information.

The Fix: Proper Scoring Rules

Utility Function Exploits IC?
Ext_Squared 5/6 No
Ext_Linear 3/6 No
Brier Score 0/6 Yes
VCG 0/6 Yes

VCG Dominates in Safety

Real NSL-KDD dataset results:

Mechanism FN Rate FP Rate
VCG 0.305 0.031
Ext_Linear 0.363 0.028
Brier Score 0.387 0.028
Majority Vote 0.380

VCG wins 7/8 FP-sensitive scenarios.

Data Efficiency: 4x Better than Neural Nets

Samples VCG (FN) DeepSets (FN)
50 0.000 0.348
100 0.000 0.270
200 0.000 0.012
500 0.001 0.004

VCG achieves perfect FN at 50 samples; DeepSets needs 200.

Communication Robustness

Failure Mode FN Rate Degradation
None 0.060
50% packet loss 0.064 +6.7%
25% corruption 0.030 -50%

Mechanism design is robust to severe communication failures — critical for edge deployment.

Practical Recommendations

  1. Safety-critical → VCG (best FN, IC)
  2. IC required, less conservative → Brier Score
  3. Low-data regime → mechanism design over neural nets (4x efficiency)
  4. Non-strategic agents → externality can work (FN=0.363)
  5. Edge deployment → mechanism design (robust to failures, scales to 200 agents at <2ms)