Timeline of Developments

A summary of major developments in CASTLE Labs over its history.

  • 1983 – Developed first interactive optimization model (“SuperSPIN”) for network design for less-than-truckload motor carriers. It was adopted by almost the entire LTL industry, and helped to stabilize LTL trucking in the post-deregulation era. SuperSPIN is still in production 40 years later (marketed by Manhattan Associates).
  • 1985-1987 – Developed MicroMAP, the first in-memory, real-time load-matching system for truckload trucking which captured the uncertainty of the future. MicroMAP was the first commercially successful load matching system, which is still in use today (marketed by Manhattan Associates).
  • 1987 – Edelman finalist with SuperSPIN implemented at Yellow Freight System.
  • 1988 – Founded Princeton Transportation Consulting Group. The initial management team was David Cape ’87 and Ken Nickerson ’84, who jointly wrote MicroMAP.  PTCG marketed SuperSPIN and MicroMAP.
  • 1990 – CASTLE Laboratory was founded with the hiring of Hugo Simao to develop an operational linehaul planning system for Yellow Freight System. First implemented in 1992, the system is still running 30 years later (with a major upgrade implemented in 2018).  CASTLE would grow to handle projects in trucking (LTL, truckload, parcel), rail (primarily locomotive optimization, management of high value spare parts, before evolving into energy, e-commerce and health.
  • 1991 – Edelman finalist for application of LOADMAP (predecessor of MicroMAP) at North American Van Lines.
  • 1992 – Developed the “LQN” methodology for fleet management, the predecessor of approximate dynamic programming for high-dimensional applications.  This methodology used a linear approximation of value functions, a technique that is still being used to optimize fleets of drivers.
  • 1994 – First real-time driver scheduling system that could assign drivers over multiple legs, all using an in-memory, real-time system (developed by Derek Gittoes).
  • 1995 – Transport Dynamics was founded by Derek Gittoes.
  • 1996-2008 – Developed PLASMA for Norfolk Southern which was the first production-quality optimization model for a North American freight railroad. The model is still running at Norfolk Southern after 15 years, and remains the only optimization model in production at a North American freight railroad.
  • 1997 – Introduced the idea of merging historical patterns in a large-scale optimization model, blending low-dimensional rules with high-dimensional algorithms.  This idea was a major breakthrough since it helped to incorporate simple patterns of behavior captured using machine learning with optimization-based methods that could handle high-dimensional applications such as managing drivers.
  • 2002 – 2008 10 years of research finally produced the first driver optimization model for truckload trucking which handles a full set of driver and load attributes, including hours of service rules, equipment types, routing drivers to home, pickup and delivery appointment windows, and home time appointments.  The model (“SMART-TL”) was adapted to Schneider National (2004-2008).  SMART-TL was later licensed to Optimal Dynamics.
  • 2002 – Greg Godfrey develops the CAVE algorithm and adapts it to resource allocation problems with multi-period travel times. This logic remains one of the core tools of the lab for a wide range of resource allocation problems.
  • 2006 – Peter Frazier introduces the “knowledge gradient,” launching a new direction in optimal learning.
  • 2007 – First edition of a book on Approximate Dynamic Programming, merging dynamic programming and math programming for the first time. The second edition (2011) was the first to identify four classes of policies.
  • 2008 – PENSA Laboratory (Princeton laboratory for ENergy Systems Analysis) was established to study stochastic optimization problems in energy, with a major grant from SAP.
  • 2009 – Winner, Daniel Wagner Prize from Informs for the first application of approximate dynamic programming for truckload fleet management for Schneider National. Now called SMART-TL, this is the first system to be able to estimate the marginal value of drivers and loads while handling all driver work rules and home constraints.  SMART-TL would be licensed to Optimal Dynamics, founded in 2017 to bring this technology to the truckload industry.
  • 2010 – Ilya Ryzhov bridges online and offline learning using the knowledge gradient.
  • 2011 – Second edition of Approximate Dynamic Programming appears, representing a major revision of the first edition. 300 new pages, and a complete restructuring of the book, including a first draft of identifying four major classes of policies.
  • 2012 – Optimal Learning is published by Wiley, introducing an entirely new class of policies for information collection tuned to the needs of business, science and engineering, geared to an undergraduate audience.
  • 2013 – “CASTLE Labs” is rechristened “Laboratory for ComputAtional STochastic optimization and LEarning” to emphasize new focus on methodology with many applications.
  • 2014 – 2019 Wrote “Jungle of Stochastic Optimization” for the Informs TutORial series.  First time the four classes of policies were presented formally for solving stochastic optimization problems.  Follow-on papers in 2016 (also for the TutORials series) and 2019 (for European J. of Operational Research) established the unified framework for sequential decision problems.  This field came to be known as “sequential decision analytics”; see the video tutorial and a webpage for educational resources.
  • 2017 – Optimal Dynamics was founded by Daniel Powell to market the “SMART” library to the truckload trucking industry. The company would later raise $50 million in funding, and by 2023 had almost 70 employees.
  • 2018 – Compiled 10+ years of research into energy storage, which is the first to fully research all four classes of policies in the context of a wide range of storage applications, including the first formal model to properly handle rolling forecasts.
  • 2022 – Published Reinforcement Learning and Stochastic Optimization: A unified framework for sequential decisions.  The first book to unify 15 fields of stochastic optimization using a single universal modeling framework that can be used to model any sequential decision problem.  The book then identifies four classes of policies that include every possible method for making decisions.
  • 2022 – Published Sequential Decision Analytics and Modeling: Modeling with Python.  This is an introductory book first used in an undergraduate course at Princeton, which uses a teach-by-example style.
  • 2023 – Posted A Modern Approach to Teaching Optimization which presents a fundamentally new approach for teaching an introduction to optimization course for undergraduates or masters.