MAFIS: Multi-Agent Fault Injection Simulator

Name: MAFIS — Multi-Agent Fault Injection Simulator
Availability: InStock
Author: Teddy Truong

Teddy Truong

MAFIS: Multi-Agent Fault Injection Simulator

A fault resilience observatory for lifelong multi-agent path finding (MAPF). MAFIS measures how multi-agent warehouse robotics systems degrade, recover, and adapt under real-world fault conditions in real-time 3D simulation.

What MAFIS Measures

Cascade failure propagation across agent networks via Action Dependency Graphs
Recovery time and throughput degradation under configurable fault injection
Heat accumulation and Weibull wear patterns on warehouse grid cells
Comparative resilience across 3 lifelong MAPF solvers (PIBT, RHCR-PBS, Token Passing)

Key Features

Real-time 3D WebAssembly simulation with up to 1000 agents
Deterministic replay via seeded randomness for reproducible experiments
Fault scenarios: burst failures, zone outages, gradual degradation
Analysis: ADG construction, cascade BFS, resilience scorecards, heatmaps
Multiple warehouse topologies with visual map editor
CSV/JSON export for offline analysis

	WASM (browser)	Desktop (native)
Expected FPS	—	—
Threading	Single-threaded	Multi-threaded
GPU backend	WebGL2	Vulkan / Metal

Metric	Formula	Range	Description
Fault Tolerance (FT)	P_faulted / P_baseline	0 – 1+	Throughput retention ratio. FT < 1 = degradation. FT > 1 = paradoxical gain (Braess: fault relieves congestion).
Critical Time (CT)	ticks_<50% / ticks_post-fault	0 – 1	Fraction of post-fault ticks where throughput falls below 50% of baseline. Lower = better.
TWTE	Σ (t − t_f) · \|1 − R(t)\|	0 – ∞	Time-weighted Throughput Error: time-weighted absolute error of the throughput ratio post-fault (Ogata 2010). Penalizes sustained degradation more heavily than early transients. Lower = better. N/A when no fault occurs.
Attack Rate (AR)	ever_affected / initial_fleet	0 – 1	Fleet-exposure fraction — proportion of agents either directly killed or transitively reached by fault cascade BFS at any point in the run (Wallinga & Lipsitch 2007). Higher = more widespread impact.
Cascade Depth	mean BFS depth_{per event}	0 – 10	Mean BFS depth on the Action Dependency Graph across fault events. Measures how far disruption propagates through agent dependencies. Capped at MAX_CASCADE_DEPTH = 10.
Rapidity	T_rec(ρ = 0.9)	ticks	Ticks from first fault until throughput recovers to ≥90% of baseline for 5 consecutive ticks (Bruneau 2003). Only applicable to recoverable faults (ZoneOutage, IntermittentFault). N/A for permanent faults. Lower = faster recovery.

Metric	Formula	Range	Description
Fault Tolerance	P_fault / P_nominal	0 – 1+	Fraction of baseline throughput retained under faults. Can exceed 1.0 (Braess paradox).
NRR	1 − MTTR / MTBF	0 – 1	Normalized Recovery Ratio. Higher = less time spent recovering. Requires ≥2 fault events and at least one cascade-affected neighbor.
Fleet Utilization	alive_tasked / initial_fleet	0 – 1	Productive fleet capacity retained. Counts latency-affected agents as tasked.
Critical Time	ticks_<50% / ticks_post-fault	0 – 1	Fraction of post-fault ticks below 50% baseline throughput. Lower is better.

Metric	Description
Throughput	Average tasks completed per tick (100-tick rolling mean).
MVA	Chart display only — 10-sample moving average applied to the throughput series for visual smoothness. Not a reported metric; the underlying 100-tick rolling mean is the actual metric.
MTTR	Mean recovery time of cascade-affected neighbors (not dead agents). 0.0 when no cascade occurs.
MTBF	Mean interval between fault events. Requires 2+ events.
Propagation Rate	Average fraction of alive fleet affected per fault event (ADG cascade-based, observatory only). Echoes fault config; Attack Rate is preferred for experiment reporting.
Survival Rate	Fraction of initial fleet still alive.
Wait Ratio	Fraction of actions where alive agents wait instead of moving. Includes solver waits, collision holds, and latency. Dead agents excluded.
Cascade Spread	Average number of agents affected per fault event via ADG BFS. Higher = wider impact per incident.
Tasks Completed	Total pickup-to-delivery tasks completed since simulation start. Compare against fault-free baseline to gauge cumulative productivity loss.
Impacted Area	Throughput deviation vs baseline: (live − baseline) / baseline × 100. Equivalent to (FT − 1) × 100. Kept for backwards compatibility.
Deficit Integral	Cumulative sum of ticks where live tasks fell behind the baseline. Higher = more total productivity lost to faults.
Surplus Integral	Cumulative sum of ticks where live tasks exceeded the baseline (Braess paradox). Net = Deficit − Surplus.

MAFIS: Multi-Agent Fault Injection Simulator

What MAFIS Measures

Key Features

Experiment Configuration

Results

WASM LIMITATIONS

DESKTOP ADVANTAGES

Fault Metrics Reference

Differential Metrics (Experiment)

Live Observatory Metrics

Composite Score

Supporting Metrics