waremax

Reading WareMax delay attribution

Every completed task decomposes into five buckets — assignment, travel, queue, congestion, service — summing to cycle time. Read it and use it as reward.

Difficulty: intermediateTime: 10 min

A single throughput number tells you nothing about where the time went. WareMax decomposes every completed task’s cycle time into five buckets that sum to the whole.

The five buckets

BucketWhat it measures
assignmentTime the task waited before a robot was assigned
travelRobot travel time to the pickup
queueTime queued at the pick station
congestionExtra time lost to traffic / wait-at-node
serviceStation service time

For a task with cycle=33.1s, a decomposition might read assign=0.2 travel=11.2 queue=5.4 service=5.5 congestion=10.8.

Why the split matters for dispatching

Only some of that time is controllable by the dispatcher. A robot’s service time at a station is not a dispatching decision; the assignment wait and travel-to-pickup are. That is exactly the split the routed reward mode uses — it charges the agent only for what it can influence, which is why it is the recommended default for RL training.

Reading it in practice

The per-task decomposition is exported alongside the event log and time-series. Aggregate it across a run to see whether your bottleneck is travel (fix routing), queue (add station concurrency), or congestion (relieve traffic) — before you blame the dispatching policy.

Related: verifying determinism · comparing dispatching policies. Built by Skelf Research.

Frequently Asked Questions

Do the five buckets always sum to cycle time?

Yes. The decomposition is a partition of the task's cycle time into assignment wait, travel, station queue, congestion, and service. It is designed to sum exactly to the observed cycle time.

How is attribution used as a reward?

The attribution reward mode uses the full per-task decomposition; the routed mode charges only the controllable part — assignment wait plus travel to pickup — which is what a dispatcher can actually influence.

Related Content

Run a deterministic experiment

Open source, MIT licensed. One YAML file and one CLI invocation.