← BONABEAU APPS

Letting the people who do the work shape how it is divided

A working model of interactive evolution for workload partitioning. The mail example follows the field process described in the Accord slides for La Poste; the scoring of the nurse example follows Funes, Bonabeau, Hervé and Morieux, “Interactive Multi-Participant Tour Allocation” (CEC 2004).

1

2

Rank 1 is the version they would most like to get. Nobody rates a whole plan.

3

The rankings choose the parents

Both goals at once

Each dot is one plan shown to participants. Lines join the plans of a round that no other plan beats on both goals, labeled with the round number. Lower left is better.

Round by round

best planmean of sixmean with no participant input

Where each plan came from

One column per round, best-ranked plan on top. Filled circles were parents of the next round.

carried overrecombinationmutation

Does workload drive the ranking?

The field process this example follows

Summarized from the Accord slides. The slides state three objectives for a reorganization: capture the appropriate level of productivity, reduce the risk of rejection at scale by winning carrier engagement, and be robust enough to be industrialized.

About one month

Preparation

Postal and map data are entered. Atomic groups are defined with carriers and first-line managers, which is where hard constraints and local knowledge enter: a dead end, a one-way street, a village that cannot be split, an uphill stretch that should stay with its downhill half.

A few hours per scenario

Design

A scenario fixes work regimes, indoor and outdoor work, and transport modes. The optimizer generates many partitions built from atomic groups, and filters remove those outside management limits.

About one week

Carrier feedback

Each morning every carrier receives six versions of their own route and ranks them. The two best partitions overall seed the next set through direct transmission, mutation and recombination.

Immediate

Vote and printing

The top three partitions go to a vote of all carriers at the office. The winner is printed as route sheets and maps.

  • What the model keeps. Atomic groups, routes in one piece, the 7h00 standard with sorting, delivery and travel time, transport modes, a management filter, daily rounds with six versions per carrier, top-two selection, the lineage of plans, and the final vote. One pair of carriers is set up to want the same two blocks with opposite priorities, as in the slide on contested groups.
  • What it leaves out. Real street and postal data, the delivery sequence inside a route, work regimes, and changes to the number of routes.
  • The reported finding. The slides show carriers’ rankings plotted against standard route length and conclude that workload does not necessarily drive satisfaction. The last chart above asks the same question of the simulated carriers.
  • Orders of magnitude. The speaker notes mention about three weeks of preparation for one office, about ten hours of design for a full repartition there, and about three hours to rebalance 30 routes at another.
How the model works and where it is simplified