Dept. of Industrial and Enterprise Systems Engineering
http://hdl.handle.net/2142/16358
Fri, 20 Apr 2018 09:16:05 GMT2018-04-20T09:16:05ZExploring image recognition: applying convoluted neural networks and learning to recognize safe cyclists
http://hdl.handle.net/2142/99141
Exploring image recognition: applying convoluted neural networks and learning to recognize safe cyclists
Narayanan, Ramakrishnan
Today, there is a need to focus on the mobility revolution that is currently taking place. With the advent of more intelligent data gathering, there is also a growing need for using existing technology and infrastructure to achieve this goal, without incorporating expensive, complicated systems. As single-occupancy give way to shared mobility solutions, combined with regular mass transit and pedestrian-aware street infrastructure (traffic lights, crosswalks etc.), there is a large "networked mobility system'' that has the potential to be tapped. Moreover, autonomous cars will be here soon, to add to the mix.
With statistics showing an increase in bicyclist related crashes over the last decade and an increase in bicycle-borne road users, there is a necessity for cities and autonomous vehicles to build bicycle safety into their adaptation to the "driverless future". This paper is an exploration into the use of a Convolutional Neural Network (CNN) based Machine Learning (ML) algorithm to identify bicycle-borne road users, who wear helmets.
We use a pre-made CNN framework-YOLO (You Only Look Once), and built around it further. After a brief proof-of-concept test on a publicly available dataset (including extraction, parsing and detection), the algorithm was modified. Some important features were added, such as identifying license plates, faces and encrypting them. Further, there is also a detailed account of using the ML capabilities that the framework is built with, and training it to identify bicycle-borne road users wearing a helmet.
Neural networks; Convolutional neural networks; Road safety; Image recognition
Fri, 21 Jul 2017 00:00:00 GMThttp://hdl.handle.net/2142/991412017-07-21T00:00:00ZNarayanan, RamakrishnanEnhancements to the perfect matching-based tree algorithm for generating architectures
http://hdl.handle.net/2142/98990
Enhancements to the perfect matching-based tree algorithm for generating architectures
Herber, Daniel Ronald; Allison, James T.
In this report, a number of enhancements to the perfect matching-based tree algorithm for generating the set of unique, feasible architectures are discussed. The original algorithm was developed to generate a set of colored graphs covering the graph structure space defined by (C,R,P) and various additional network structure constraints. The proposed enhancements either more efficiently cover the same graph structure space or allow additional network structure constraints to be defined. The seven enhancements in this report are replicate ordering, avoiding loops, avoiding multi-edges, avoiding line-connectivity constraints, checking for saturated subgraphs, enumerating subcatalogs, and alternative tree traversal strategies. Some theory, implementation details, and examples are provided for each enhancement.
algorithm; architecture; graphs; perfect matchings; matlab
Sun, 10 Dec 2017 00:00:00 GMThttp://hdl.handle.net/2142/989902017-12-10T00:00:00ZHerber, Daniel RonaldAllison, James T.Data driven approaches to improve operational efficiency of emergency medical services
http://hdl.handle.net/2142/98408
Data driven approaches to improve operational efficiency of emergency medical services
Krishnan, Kaushik
We study data-driven approaches to maximize the service level of Emergency Medical Services (EMS) in emerging economies. These systems usually operate under heavy resource constraints and face significant operational challenges, making them structurally and operationally different from systems in developed countries. In this thesis we study two specific issues - (i) modeling human behavior, and (ii) accounting for risk metrics due to tail behavior.
First, we address the issue of ambulance abandonment that occurs when a patient's willingness to wait is less than the ambulance response time resulting in the vehicle not being utilized. We present a maximum likelihood estimation approach to estimate willingness to wait for different types of patients. We then use the estimate of waiting times in a greedy simulation based optimization model to redesign the EMS network to maximize the number of patients served within their waiting time thresholds. Computational experiments using data from an Indian metropolitan city show that our proposed resource allocation model reduces abandonment by approximately 2 percentage points with the current ambulance fleet size, 5 percentage points by doubling the fleet size and 6 percentage points by tripling the fleet size.
Next, we present a risk-based optimization approach to make the EMS network robust to unexpected changes in demand patterns. This is motivated by the fact that when few parts of the network face heavy-tailed demand patterns, the demand for entire network under the resource constrained setting behaves in a heavy-tailed manner. To achieve a robust location strategy we include risk metrics, specifically the Conditional Value at Risk, that focus on tail behavior in addition to average case performance metrics. Computational experiments show that planning with a view of minimizing risk leads to solutions that perform well in heavy-tailed settings.
Maximum likelihood estimation; Conditional value at risk
Mon, 17 Jul 2017 00:00:00 GMThttp://hdl.handle.net/2142/984082017-07-17T00:00:00ZKrishnan, KaushikAlternative design optimization formulations – developing and comparing for a vibration damping example
http://hdl.handle.net/2142/98306
Alternative design optimization formulations – developing and comparing for a vibration damping example
Luan, Siyao
Developing mathematical formulations for design problems requires determining an objective function to compare design alternatives, a set of design features and options to be included for consideration, and a predictive model that reflect any unavoidable cause and effect relationships that are relevant. There are no formal principles guiding the formulation process, and heuristics prevail. There are some instances when the level of effort required to formulate the problem and solve the problem are excessive, not worth the improvement that might be realized in the overall design objective. In other words, a hypothetically “perfect” design problem formulation that takes all possible factors into account might be so difficult to fully compose and solve that it is not worth the effort. This thesis presents a set of guidelines for formulating design problems that seeks a middle ground. The method presented defines three different tasks in the formulation process: comparison metrics, predictive model and design representation. Each task offers opportunities for the practitioner to balance the expected quality of the solution with the level of effort and time required to reach that solution. This thesis demonstrates how using the guidelines can help create alternative formulations for the same design problem, and then how the resulting solutions can be evaluated and compared. Using a vibration absorber design example, the guidelines are enumerated, explained, and used to compose six alternative optimization formulations of the problem. These alternatives formulations vary in objective functions, decision variables, and some other design formulation practices. The overall goal is to maximize surface finish quality of a machined component processed on a platform to which the vibration absorber is attached. Vibrations of the platform can have a detrimental effect on surface quality. The goal of the vibration absorber system is to minimize these detrimental effects. The six alternative optimization formulations are subsequently solved, and their scores reflecting their complexity, computational time and solution quality are quantified and compared. The results illustrate the unavoidable tradeoffs among these three attributes. The best formulation depends on the set of tradeoffs that are best in that situation, given the decision maker’s risk attitude and preference.
Optimization formulation; Design guidelines
Thu, 20 Jul 2017 00:00:00 GMThttp://hdl.handle.net/2142/983062017-07-20T00:00:00ZLuan, SiyaoCausal structure of networks of stochastic processes
http://hdl.handle.net/2142/98151
Causal structure of networks of stochastic processes
Etesami, Seyedjalal
We propose different approaches to infer causal influences between agents in a network using only observed time series. This includes graphical models to depict causal relationships in the network, algorithms to identify the graphs in different scenarios and when only a subset of agents are observed. We demonstrate the utility of the methods by identifying causal influences between markets and causal flow of information between media sites.
We study the statistical and functional dependencies in network of processes. Statistical dependencies can be encoded by directed information graphs (DIGs) and functional relationships using functional dependency graphs (FDGs), both of which are graphical models where nodes represent random processes. DIGs are based on directed information that is an information theoretic quantity. To capture the functional dependencies in a dynamical system, we introduce a new measure in this work and show that the FDGs are a generalization of DIGs. We also establish sufficient conditions under which the FDG defined by our measure is equivalent to the DIG. As an example, we study the relationship between DIGs and linear dynamical graphs (LDGs), that are also a type of graphical models to encode functional dependencies in linear dynamical systems. In this case, we show that any causal LDGs can be reconstructed through learning the corresponding DIGs.
Another contribution is to propose an approach for learning causal interaction network of mutually exciting linear Hawkes processes. In such processes, a natural notion of functional causality exists between processes that is encoded in their corresponding excitation matrices. We show that such causal interaction network is equivalent to the DIG of the processes. Furthermore, We present an algorithm for learning the support of excitation matrix (or equivalently the DIG). The performance of the algorithm is evaluated for a synthesized multivariate Hawkes network as well as real world dataset.
We also study the problem of causal discovery in presence of latent variables, in which only a subset of processes can be observed. We propose an approach for learning latent directed polytrees as long as there exists an appropriately defined discrepancy measure between the observed nodes. Specifically, we use our approach for learning directed information polytrees. We prove that the approach is consistent for learning minimal latent directed trees. Furthermore, we study the problem of structural learning in vector autoregressive (VAR) models with latent variables. In this case, we extend the identifiability to a broader class of structures. In particular, we show that most of the causal structure of a VAR model can be recovered successfully when only the causal network among the latent variables is a directed tree.
Last but not least, we introduce a new statistical metric inspired by Dobrushin’s coefficient [1] to measure the dependency or causal direction between variables from observational or interventional data. Our metric has been developed based on the paradigm that the conditional distribution of the variable of interest given all the direct causes will not change by intervening on other variables in the system. We show the advantageous of our measure over other dependency measures in the literature.
We demonstrate the effectiveness of the proposed algorithms through simulations and data analysis.
Causal learning; Dynamical systems; Stochastic systems
Wed, 31 May 2017 00:00:00 GMThttp://hdl.handle.net/2142/981512017-05-31T00:00:00ZEtesami, SeyedjalalStochastic optimization with decisions truncated by random variables and its applications in operations
http://hdl.handle.net/2142/98234
Stochastic optimization with decisions truncated by random variables and its applications in operations
Gao, Xiangyu
We study stochastic optimization problems with decisions truncated by random variables and its applications in operations management. The technical difficulty of these problems is that the optimization problem is not convex due to the truncation. We develop a transformation technique to convert the original non-convex optimization problems to convex ones while preservation some desired structural properties, which are useful for characterizing optimal decision policies and conducting comparative statics. Our transformation technique provides a unified approach to analyze a broad class of models in inventory control and revenue management. In additional, we develop efficient algorithms to solve the transformed stochastic optimization problem.
Stochastic optimization; Convexity; Supply capacity uncertainty; Revenue management
Mon, 26 Jun 2017 00:00:00 GMThttp://hdl.handle.net/2142/982342017-06-26T00:00:00ZGao, XiangyuGPU accelerated Hungarian algorithm for traveling salesman problem
http://hdl.handle.net/2142/97805
GPU accelerated Hungarian algorithm for traveling salesman problem
Kaushik, Varsha Ravi Prakash
In this thesis, we present a model of the Traveling Salesman Problem (TSP) cast in a quadratic assignment problem framework with linearized objective function and constraints. This is referred to as Reformulation Linearization Technique at Level 2 (or RLT2). We apply dual ascent procedure for obtaining lower bounds that employs Linear Assignment Problem (LAP) solver recently developed by Date(2016). The solver is a parallelized Hungarian Algorithm that uses Compute Unified Device Architecture (CUDA) enabled NVIDIA Graphics Processing Units (GPU) as the parallel programming architecture. The aim of this thesis is to make use of a modified version of the Dual Ascent-LAP solver to solve the TSP.
Though this procedure is computational expensive, the bounds obtained are tight and our experimental results confirm that the gap is within 2% for most problems. However, due to limitations in computational resources, we could only test problem sizes N < 30. Further work can be directed at theoretical and computational analysis to test the efficiency of our approach for larger problem instances.
Compute Unified Device Architecture (CUDA); Linear assignment problem; Traveling salesman problem; Reformulation Linearization Technique (RLT)
Fri, 28 Apr 2017 00:00:00 GMThttp://hdl.handle.net/2142/978052017-04-28T00:00:00ZKaushik, Varsha Ravi PrakashOptimization and control of wind energy systems for grid integration
http://hdl.handle.net/2142/97653
Optimization and control of wind energy systems for grid integration
Deshmukh, Anand Pramod
Wind energy is a rapidly expanding source of renewable energy, but wind resources are highly intermittent. This makes increasing the level of wind energy penetration in an overall energy portfolio challenging if power quality and grid frequency is to be maintained. In a conventional power system, grid frequency regulation is typically achieved by means of some form of active power control (APC) of power generation plants. Active power control of plant power output aims to maintain the power balance between generation and consumption. Wind turbines have historically not participated in the active power control and are therefore isolated from the grid using sophisticated power electronics, increasing the cost of wind energy. Interest in studying APC of wind turbines for grid frequency regulation has been revived recently. Most of the proposed approaches either focus on single turbine, or overlook the effect of APC strategies on actuator usage and mechanical loading of the system. However, wind energy based power generation plants have an array of wind turbines that interact with each other aerodynamically in a complicated manner. In this work we introduce a new distributed APC strategy in which a farm level controller optimally distributes the task of regulation to all the wind turbines in a farm accounting for dynamic wake effects introduced due to control actions of each of those wind turbines. An individual model predictive controller at each wind turbine then tracks the power references passed on by farm level controller, subject to mechanical loading constraints. The results from this approach are compared with the greedy approach when the individual wind turbines only optimize their own power production without consideration of downstream neighbors. We then extend the idea of this hierarchical control to co-optimization of wind farms and battery energy storage for regulation, and then for simultaneous wind farm layout and control design to exploit the synergy between layout design and wind farm control.
We show that the regulation performance scores for the distributed APC approach are statistically significantly better than the greedy approach. We also show that co-optimization of wind farm and battery energy storage provides significantly superior performance over baseline, for a much smaller battery capacity, providing a potentially significant cost saving. Finally, we show that designing a layout with simultaneous consideration of control system design, allows us harness synergistic relationships between layout and control. Capitalizing on these synergy mechanisms enables increased annualized wind farm energy production.
Wind energy; Model predictive control; Automatic generation control
Thu, 02 Feb 2017 00:00:00 GMThttp://hdl.handle.net/2142/976532017-02-02T00:00:00ZDeshmukh, Anand PramodAnalysis of a process and computer user interface for data analytics results for patient discharge decision making
http://hdl.handle.net/2142/97397
Analysis of a process and computer user interface for data analytics results for patient discharge decision making
Jain, Sanjana
Initially, computer aids simply automated the processes that humans had developed to complete various tasks. Their primary contribution was to increase efficiency by automating repetitive, tedious tasks, both physical and cognitive. Many years later, sophisticated computer aids employ a broad-scale availability of large amounts of information, coupled with the ability to analyze that information using statistical methods. Data analytics can provide more information than ever before to those using computer based systems. Computer aids for engineering design are typically used by engineers, who are data-savvy, and are generally capable of using the results of data analytics directly incorporated into engineering design tools in an efficient manner. However, data analytics results are becoming more widespread and available for users who are not data-savvy. These systems must be designed with this in mind. This paper presents an example of a computer aid for healthcare, a tool to aid in the patient discharge decision. The motivating problem is the large number of patients who are readmitted to hospitals within 30 days of discharge. The design problem addressed here includes identifying the decision process the users should employ, as well as the computer/user interface that best uses the data analytics results. This work was carried out through collaboration with a healthcare provider. A readmission risk tool was developed using data analytics to estimate the probability of readmission based on a historical data set of 50 patient-specific factors. Preliminary work indicated that the users did not fully accept or utilize the data analytics results, and also demonstrated their need to continue to rely in part on their own expert heuristic decision processes. This study presents a method for dealing with these issues. First, a normative decision based approach to determining a multiattribute utility function is formulated and assessed to compare alternative computer/user interface designs. Then, the same approach is taken to assess the healthcare worker’s willingness to make tradeoffs under uncertainty when making the patient discharge decision. The resulting interactive computer interface design coupled with the normative healthcare decision process helps the user best exploit data analytics results while simultaneously facilitating the user’s continued deployment of their own expertise.
Currently, the RRT is being utilized for flagging the readmission risk prone patients and not for making the main discharge decision regarding the patient. Its main use is to help the healthcare workers focus their attention on these medium-high risk patient profiles. Furthermore, the RRT result display needs to be made more actionable and relevant to the healthcare workers for the efficient use of the RRT model. We will discuss this in detail in our study.
Readmission risk tool (RRT); Patient discharge; Decision; Healthcare; Risk tool; Design
Mon, 24 Apr 2017 00:00:00 GMThttp://hdl.handle.net/2142/973972017-04-24T00:00:00ZJain, SanjanaGeneralized sequential assignment problem
http://hdl.handle.net/2142/97329
Generalized sequential assignment problem
Khatibi, Arash
The Sequential Stochastic Assignment Problem (SSAP) deals with assigning sequentially arriving tasks with stochastic parameters to workers with fixed success rates. The reward of each assignment is the product of the worker's success rate and the task value assigned to the worker. The objective is to maximize the total expected reward. There has been a surge of interest in studying sequential assignment problems due to their applications in online matching markets, asset selling, and organ transplant.
This dissertation studies several variations of SSAP by relaxing the main assumptions. The first part assumes that the workers' success rates are random values coming from a known distribution. This generalization modifies the SSAP from a problem with a single random value (i.e., the task value) at each stage to an online matching problem with several random parameters (i.e., the task value and the workers' success rates). The optimal assignment policy uses backward induction to first solve smaller subproblems, and then use them to optimally assign tasks to workers from the first stage. An approximation algorithm is proposed that achieves a fraction of the optimal reward in a polynomial time.
Assuming that the value of sequentially arriving elements are independently drawn from a known distribution is unrealistic in many applications. The second part of thesis relaxes this assumption and uses the well-known Secretary Problem to derive constant-competitive algorithms for SSAP with tasks having a random arrival order. Several deterministic and randomized algorithms are proposed and their performance are compared with the maximum offline reward. These algorithms use the first stages of the problem as a training phase to compute thresholds for the task values. These thresholds are used to assign tasks to workers after the training phase.
The third part uses the linear programming technique to derive bounds on the performance of optimal policy for several variations of SSAP. Formulating an online matching problem as a linear program is a useful tool. In addition to deriving bounds on performance of optimal policies, the linear programming technique can be used to formulate extensions of the problem as linear programs by simple changes in the objective function and constraints of the basic formulation. The linear programming formulation of the incentive compatible problem and the sequential assignment problem with unknown number of elements are also proposed. The edge-weighted online bipartite matching problem is used to design assignment policies for each of the formulated problems.
The last part relaxes the assumption that at most one task must be assigned to each worker in SSAP. It is assumed that a worker is available for possible future assignments after performing the previously assigned task. The number of stages that the worker is not available due to a prior task assignment is referred to as the task duration. This problem is studied under various models for the task duration. First, it is assumed that the task duration is fixed. Then, assignment policies are proposed for the problem with a memoryless model for the task duration. The proposed algorithms are extensions of the optimal algorithm for the sequential assignment problem. They divide the n-stage assignment process to periods whose lengths are equal to the expected task duration. Then, they assign tasks to workers in each period by applying the optimal algorithm of the sequential assignment problem.
Linear program; Online matching; Secretary problem; Sequential assignment
Tue, 11 Apr 2017 00:00:00 GMThttp://hdl.handle.net/2142/973292017-04-11T00:00:00ZKhatibi, Arash