No such thing as a perfect hammer: comparing different objective function specifications for optimal control

Abstract

Linear-quadratic (LQ) optimization is a fairly standard technique in the optimal control framework. LQ is very well researched, and there are many extensions for more sophisticated scenarios like nonlinear models. Conventionally, the quadratic objective function is taken as a prerequisite for calculating derivative-based solutions of optimal control problems. However, it is not clear whether this framework is as universal as it is considered to be. In particular, we address the question whether the objective function specification and the corresponding penalties applied are well suited in case of a large exogenous shock an economy can experience because of, e.g., the European debt crisis. While one can still efficiently minimize quadratic deviations around policy targets, the economy itself has to go through a period of turbulence with economic indicators, such as unemployment, inflation or public debt, changing considerably over time. We test four alternative designs of the objective function: a least median of squares based approach, absolute deviations, cubic and quartic objective IS acknowledges financial support from the German Science Foundation (DFG RTG 1411), the Helmholtz Association (HIRG-0069) and the Projex CSES, Initiative d'Excellence, Université de Strasbourg. We also thank anonymous referees for very helpful comments and suggestions. The usual disclaimers apply. functions. The analysis is performed based on a small-scale model of the Austrian economy and illustrates a certain trade-off between quickly finding an optimal solution using the LQ technique (reaching defined policy targets) and accounting for alternative objectives, such as limiting volatility in economic performance. As an implication, we argue in favor of the considerably more flexible optimization technique based on heuristic methods (such as Differential Evolution), which allows one to minimize various loss function specifications, but also takes additional constraints into account.

An Evolutionary Approach to Passive Learning in Optimal Control Problems

“…As a small open economy, Austria's growth rate of GDP heavily depends on the economic situation in the world via exports. In order to hold the model system close to reality, this is captured using the correlation coefficient (a 1 ) between Austria's and world 's GDP growth (between 1996 and2017) and is equal to 0.7266. In addition, the government is able to use its fiscal policy instrument (g t ), which is primary fiscal surplus (or deficit if negative).…”

An Evolutionary Approach to Passive Learning in Optimal Control Problems

“…In particular, the restriction of the symmetric objective function are addressed in Blueschke et al (2013b), while inequality constraints-in Savin et al (2018). In addition, Blueschke and Savin (2017) compare alternative forms of objective function specification, which may be more suitable if-next to achieving policy targets-one is also willing to minimize volatility of the economic system over time.…”

“…( 2018 ). In addition, Blueschke and Savin ( 2017 ) compare alternative forms of objective function specification, which may be more suitable if—next to achieving policy targets—one is also willing to minimize volatility of the economic system over time.…”

Deep Learning Networks with p-norm Loss Layers for Spatial Resolution Enhancement of 3D Medical Images

“…In recent times, the use of p-norm methods in optimization algorithms has drawn attention for various applications. They have been used for machine learning binary classifiers [5,4], and in the framework of optimal control [2]. Due to the properties of the p-norm when dealing with with sparse vectors and matrices [3,8,14,1], it has also been applied for feature selection [13].…”