Events

Who: Arik Senderovich, Lyon Sachs postdoctoral fellow, University of Toronto   When: Thursday, June 21st @ 12:00pm – 1:00pm Where: BA8256

Abstract:  Process mining is a rapidly evolving research field that aims at discovering process models, such as queueing networks and stochastic Petri nets, from transactional data. On the one hand, process mining creates models that can be used for operational analysis (e.g. staffing and wait time estimation). On the other hand, operations research methods can be used to improve process discovery. This talk will mainly focus on the inter-relations between process mining and operations research through the story of automated model simplification. We shall demonstrate how queueing theory and combinatorial optimization can be applied in order to improve the quality of discovered process models.

Who: Hassan Anis, M.A.Sc Candidate, University of Toronto

When: Wednesday, December 6th @ 12:00pm – 1:00pm

Where: RS207

Abstract: One of the most important problems faced by stock traders is how to execute large block orders of security shares. When liquidating a large position, the trader faces the following dilemma: a slow trading rate risks prices moving away from their current quote, while a faster trading rate will drive quotes away from the current one leading to a large market impact. We propose a novel quasi-multi-period model for optimal position liquidation in the presence of both temporary and permanent market impact. Four features distinguish the proposed approach from alternatives. First, instead of the common stylized approach of modelling the problem as a dynamic program with static trading rates, we frame the problem as a stochastic SOCP which uses a collection of sample paths to represent possible future realizations of state variables. This, in turn, is used to construct trading strategies that differentiate decisions with respect to the observed market conditions. Second, our trading horizon is a single day divided into multiple intraday periods allowing us to take advantage of the seasonal intraday patterns in the optimization. This paper is the first to apply Engle’s Multiplicative component GARCH to estimate and update intraday volatilities in a trading strategy. Third, we implement a shrinking horizon framework to update intraday parameters by incorporating new incoming information while maintaining standard non-anticipativity constraints. We construct a model where the trader uses information from observations of price evolution during the day to continuously update the size of future trade orders. Thus, the trader is able to dynamically update the trading decisions based on changing market conditions. Finally, we use asymmetric measures of risk which, unlike symmetric measures such as variance, capture the fact that investors are usually not averse to deviations from the expected target if these deviations are in their advantage.

Who: Buser Say, Ph.D. Candidate, University of Toronto

When: Wednesday, November 22nd @ 12:00pm – 1:00pm

Where: BA3008

Abstract: In many real-world hybrid (mixed discrete continuous) planning problems such as Reservoir Control, Heating, Ventilation and Air Conditioning (HVAC), and Navigation, it is difficult to obtain a model of the complex nonlinear dynamics that govern state evolution. However, the ubiquity of modern sensors allow us to collect large quantities of data from each of these complex systems and build accurate, nonlinear deep network models of their state transitions. But there remains one major problem for the task of control – how can we plan with deep net- work learned transition models without resorting to Monte Carlo Tree Search and other black-box transition model techniques that ignore model structure and do not easily extend to mixed discrete and continuous domains? In this paper, we make the critical observation that the popular Rectified Linear Unit (ReLU) transfer function for deep networks not only allows accurate nonlinear deep net model learning, but also permits a direct compilation of the deep network transition model to a Mixed- Integer Linear Program (MILP) encoding in a planner we call Hybrid Deep MILP Planning (HD-MILP-PLAN). We identify deep net specific optimizations and a simple sparsification method for HD-MILP-PLAN that improve performance over a naive encoding, and show that we are able to plan optimally with respect to the learned deep network.

 

 

 

Who: Carloyn Busby

When: Tuesday, November 7th @ 12:00pm – 1:00pm

Where: RS207

Abstract: Many Canadian hospitals run at or near capacity, frequently experiencing congestion due to surges in demand. “Surge Protocols” that formally define when and what kind of operational steps can be taken to alleviate congestion are routinely in use. Decisions across the hospital, regarding bed capacity and allocation, staffing levels, and the surgical block schedule influence the frequency and severity of congestion, which in turn manifests in high bed occupancy, delayed admissions, a crowded Emergency Department, surgical cancellations and increased use of surge protocols. A generic, data-driven, discrete event simulation is presented that helps hospitals assess the impact of hospital wide decisions and surge policies on each area of the hospital. The model was developed in cooperation with two hospitals, and then applied at two additional hospitals.

Who: Justin Boutilier, Ph.D. Candidate, University of Toronto

When: Wednesday, October 11th @ 12:00pm – 1:00pm

Where: RS207

Abstract: Time sensitive medical emergencies are a major health concern comprising one third of all deaths in low and middle income countries (LMICs). Despite evidence that emergency transport services can save lives, poor access and availability of emergency medical care in LMICs continues to be a widespread problem. In this paper, we develop a two-stage robust optimization model that optimizes both the location and routing of emergency response vehicles, incorporating both uncertain travel times and uncertain spatial demand, which are derived from real LMIC data.