日時: |
5月12日(木) 13:30~14:30 |
会場: |
情報科学研究科A棟 A210/A212
|
内容: |
博士後期課程中間発表会 |
|
講演者: |
ZHANG ZHICHENG(藤崎研) |
講演題目: |
Probabilistic Guarantees in Sparse Robust Optimization: Theory and Applications |
概要: |
Sparse robust optimization is a general sparse optimization with uncertain constraints that minimizes the decision variables associated with L_0 cost to promote the sparsity as well as hedges against the uncertainty to satisfy the robustness.
A crucial step for robust analysis of uncertainty in this study is probabilistic relaxation on feasibility, that is, uncertain variable is modeled as a finite number of i.i.d. randomly sampled "scenarios" from a probability distribution.
Besides, the challenge of sparse optimization is to perform the L_0 norm program, namely, calculate the support or cardinality of decision variable set, which is an non-convex optimization and hence, it is extremely difficult to solve.
In this study, we introduce convex relaxation in cost function and shows that the relaxed L_1 norm still induces the sparse decision.
Therefore, the original sparse robust optimization converts into a convex L_1 norm scenario optimization problem, and the obtained sparse scenario solution is with a high probabilistic robustness for uncertainty.
Finally, we illustrate the theoretical results on sparse robust control in uncertain discrete-time systems.
Numerical results suggest that the proposed method provides a high probabilistic robustness guarantee for the obtained sparse control input to practical control system applications.
|
|
講演者: |
ASAVANANT TISSAWAT (森田研) |
講演題目: |
Forecastability analysis for passenger rail network Origin-Destination matrix: Statistically distributed variability approach |
概要: |
Static OD-matrix Estimation problem gives seed matrix representation of the demand which is then uses as input for dynamic extension given correlation to additional information.
In rail network, seed matrix statistically represents the average value explaining the variability of the demand and is the deciding factor for scheduling problem.
Statistical analyses and descriptive characterizations are sometimes assumed to be offering information on time series forecastability.
Despite the scientific interest of multiple models using varieties of data, the relationships between features (e.g, temporal dependence or time slices, trend, linearity features) and actual variability (evaluation of estimated seed matrix and actual demand) are scarcely analyzed in the literature.
In this study, multiple observations of entering time-based OD-matrices of rail networks are used to developed systematic distribution forecastability analysis framework to fill the gap in literature.
First, a multinomial distributed data is assumed, and parameter estimation is conducted using maximum likelihood estimator (MLE) from multivariate discrete data and multivariate transformed sampling proportions.
Secondly, linearly defined homogenous scaling factor is used to classify ODM components by critical condition derived from related distributions attributing to the evaluation of the seed matrix.
Lastly, forecastability evaluation criteria are investigated by sub-components of the OD-matrix.
The proposed method was tested using real data of 2 subway networks (P002 and P502) of Bangkok, Thailand.
Our proposed framework demonstrated the forecastability of the network depending on factors (i.e., time-slices and calendar classification, users' behavior analysis and stability, proportionality approach and individual discrete bin, and classification-based forecasting criteria).
|