Modeling Supply Chain Lag

Inspired by the Beer Distribution Game — created at MIT Sloan in the early 1960s to teach the core dynamics of supply chain management — Drivetrain re-tells the same story with bikes.

A small drivetrain of nodes moves product downstream: two suppliers (one for frames, one for component kits of wheels and derailleurs) feed a single central warehouse, which assembles finished bikes and distributes them to three stores. Only the stores see real customer demand.

In this lab we explore the bullwhip effect — how small fluctuations in demand at the store level get amplified into progressively larger swings upstream at the warehouse and suppliers.

Active R&D Lab — Under Construction

Multi-Agent Supply Chain Simulation

We are building an interactive 4-stage supply chain simulator using multi-agent reinforcement learning. This lab will demonstrate how lead-time lag, holding costs, and backorder penalties propagate downstream to produce the classic bullwhip effect, and how AI-driven coordination policies can mitigate it.

Simulation Node Topology

Retailer

Downstream demand

➔

Wholesaler

Local hub

➔

Distributor

Regional logistics

➔

Manufacturer

Production line

Implementation Backlog

⚙

Interactive replenishment control

Allow visitors to manually input orders for one or more supply chain nodes.

⚙

DQN and PPO reinforcement learning agents

Pre-trained agents playing the game with optimal policy weights.

⚙

Live bullwhip plotting with D3

Real-time charts rendering order amplification, inventory levels, and backlogs.

⚙

Dynamic parameter overrides

Configurable lead times, information sharing latency, and demand patterns.

Why does this happen?

The primary drivers of the bullwhip effect are:

Lack of transparency: Each node only sees orders from its immediate downstream neighbour, not true end-customer demand.
Lead times: The delay between placing an order and receiving goods causes over-ordering, especially when production and shipping stack up.
Batch ordering: Ordering in large batches to save on transport costs distorts the demand signal.

Try pushing the supplier lead time higher while adding demand variability: the supplier-order line in the amplification chart swings far wider than the underlying customer demand, and the bullwhip multiplier climbs above one. By modelling these systems with reinforcement learning and operations-research techniques, we can design control policies that drastically mitigate this effect.