Adrien Bolling

Interactive · after the KATA and ONGOING papers

Dispatch

Ten days, five machines, four technicians. Machines fail, and you choose who repairs them. Marta, your veteran, retires at the start of day 5. Can you keep the plant running after she leaves?

Time Day 1 · 00:00
Units made 0
Crew knowledge 100
Season

Machines

    Waiting tickets 0

      No ticket waits. All repairs are assigned.

      Log

        Technicians

        Each grid is a knowledge map over the task space, on the novice-to-expert scale below. The ring marks the selected ticket, and the labels mark the tasks of each machine.

          How to play

          1. A machine fails, and a ticket appears. The machine makes nothing until the repair is done.
          2. Select the ticket, then one technician. Select two technicians to pair a mentor with a learner. Press Dispatch.
          3. Experience makes repairs faster, and a good repair lasts longer. The grids show what each technician knows.
          4. Work makes technicians tired. A tired technician is slower. An exhausted technician takes 24 h of sick leave.
          5. Marta retires at the start of day 5. Her knowledge leaves with her, unless someone learned it.

          Keys: Space run or pause · 1–4 select a technician · Enter dispatch

          How it works

          Dispatch is a small version of two models from my research. The model is a few hundred lines of JavaScript, and it runs in your browser. I tuned the constants by hand, so that the effect is visible in ten days.

          Knowledge grids

          From ONGOING. Each technician has a 12 × 12 grid over the task space. Tickets from one machine land in one region. A repair adds a Gaussian bump of experience e around its cell, so related tasks gain a part of it.

          Learning curve

          The repair time is t = tnovice · (1 + e)−b · exp(0.5 f), with b = −log₂ 0.8 ≈ 0.32. This is Wright's law: each doubling of experience cuts the time by 20 %.

          Mentoring

          Two technicians on one ticket work 15 % faster. Near the ticket, the learner receives half of the difference between the mentor's experience and their own. This is the transmission mechanism of ONGOING.

          Quality and fatigue

          A repair by a novice lasts less long: the machine runs only 35 % of the normal time until its next failure. Fatigue f grows by 0.06 per hour of work and falls by about 10 % per hour of rest. At 1.0, the technician takes 24 h of leave.

          The three policies

          Fastest technician optimizes production only. Human-centric rule also rests tired technicians and pairs a learner with a mentor when no other ticket waits. Random picks any free technician. All three see the same failures as you.

          What the paper does

          The KATA study uses the full FactoReal simulator and learns a policy with a Transformer (HTT-RL), not hand-written rules. Over 200 seasons of this game, the human-centric rule beats the fastest-technician policy in 191.