## Friday, September 17, 2021

### Reallocate people: a very small but interesting LP

This is a strange little problem [1]. We have individuals (of different categories) and regions with a given capacity. The initial set of assigned persons fits within the capacity of each region. Now a new batch of individuals arrives with their chosen zones. Unfortunately, the total demand (old + new individuals) does not fit anymore in all cases. This means we need to reallocate people. We can do this at a price (the costs are given). What is the optimal (least cost) new allocation?

## Friday, September 10, 2021

### Huber regression: different formulations

 Peter Huber, Swiss statistician

Traditionally, regression, well-known and widely used, is based on a least-squares objective: we have quadratic costs. A disadvantage of this approach is that large deviations (say due to outliers) get a very large penalty. As a result, these outliers can affect the results dramatically.

One alternative regression method that tries to remedy this is L1- or LAD-regression (LAD=Least Absolute Deviation) [1]. Here I want to discuss a method that is somewhere in the middle between L1 regression and least-squares: Huber or M-Regression [2].  Huber regression uses a loss function that is quadratic for small deviations and linear for large ones. It is defined in a piecewise fashion. The regression can look like the following non-linear optimization problem:

NLP Model
\begin{align} \min_{\color{darkred}\beta,\color{darkred}\varepsilon} &\sum_i \rho(\color{darkred}\varepsilon_i) \\ &\color{darkblue}y_i = \sum_j \color{darkblue}x_{i,j}\color{darkred}\beta_j +\color{darkred}\varepsilon_i \end{align} where $\rho(\color{darkred}\varepsilon_i) = \begin{cases}\color{darkred}\varepsilon^2_i & \text{if |\color{darkred}\varepsilon_i | \le \color{darkblue}k} \\ 2\color{darkblue}k |\color{darkred}\varepsilon_i |-\color{darkblue}k^2 & \text{otherwise}\end{cases}$