Simulation Design

Purpose

Generate data for paper Sections 7.1–7.4 (convergence, shock, Sybil, EWMA sweep).

Agent-Based Model (Python)

Environment

• 50 sinks, 10 sources, 3 task types
• Discrete time steps (1 step = 1 block ≈ 2 seconds on Base)
• Each sink has internal capacity drawn from distribution (heterogeneous)

Agent Behaviors

• Source: emits payment at constant or bursty rate
• Sink: signals capacity with noise (honest ± ε), processes work, earns payment
• Malicious: over-reports capacity, under-delivers (for Sybil/slashing experiments)

Metrics

1. Allocation efficiency: Σ min(payment_received, actual_capacity) / total_payment - measures how much payment goes to productive use
2. Convergence time: steps until allocation is within 5% of optimal
3. Sybil profit: gain from splitting vs honest single-sink
4. Queue stability: max virtual queue backlog over time

Experiments

#	Name	Variables	Expected Result
E1	Convergence	BPE vs round-robin vs random	BPE converges to capacity-proportional
E2	Shock	Kill 20% sinks at t=500	BPE rebalances in O(1/α) steps
E3	Sybil	k = {1,2,5,10,20} fragments	Net profit decreasing in k
E4	EWMA	α =	0.2-0.3 optimal stability/responsiveness
E5	Buffer	B_max = {0, 0.1, 0.5, 1.0} × flow_rate	Buffer prevents source stalling

Implementation

# Core simulation loop (pseudocode)
for t in range(T):
    for source in sources:
        source.emit_payment(t)
    for sink in sinks:
        sink.signal_capacity(t)  # + EWMA
    backpressure_allocate(sinks, sources)  # max-weight
    for sink in sinks:
        sink.process_and_earn(t)
    collect_metrics(t)

Tech

• Python 3.11+, NumPy, Matplotlib/Seaborn
• Results: CSV + plots (PDF for paper)