Adaptive Dynamic Programming (ADP)

UT Arlington Faculty: Victoria Chen, Bill Corley, Jay Rosenberger
Collaborators: Cristiano Cervellera, Seoung Bum Kim
Ph.D. Students: Julia Tsai (Georgia Tech 2002), Zehua Yang (2004), Aihong Wen (2005), Dachuan Shih (2006), Huiyuan Fan (2008), Prashant Tarun (2008), Bancha Ariyajunya, Diana Martinez, Surachai Charoensri, Subrat Sahu (2011), Bancha Ariyajunya (2012), Diana Martinez (2013), Asama Kulvanitchaiyanunt (2014), Zirun Zhang (2015), Ying Chen (2017), Ashkan Aliabadi Farahani (2019), Xinglong Ju (2019)

Funding: National Science Foundation 2008-11, 2011-15
Topics: Design and analysis of computer experiments (DACE) based ADP, Sequential algorithms, State transition modeling, Value function approximation

Description: Many problems require that decisions be made over time, making dynamic programming models particularly applicable. Examples include control of inventory systems, revenue management, mechanical operations, environmental systems, behavioral intervention programs, and adaptive medical treatment. Because exact solutions for these highly complex systems are computationally intractable, we have developed methods yielding approximate solutions to some of the largest problems in the literature. The COSMOS approach involves a statistical perspective using experimental design, computer experiments, and data mining to generate parsimonious approximations that model only the critical information. COSMOS applications that involve dynamic decision-making include: airline revenue management, joint with Sabre Holdings; environmentally-friendly airplane deicing decisions, joint with Dallas-Fort Worth International Airport; assignment of nurses to patients admitted during a shift, and ozone pollution control strategies. In particular, our methods are being adapted to handle multiple objectives, convexity vs. nonconvexity, and a mix of continuous and categorical variables.