Feature Article: Optimization for simulation: Theory vs. Practice

Abstract

Probably one of the most successful interfaces between operations research and computer science has been the development of discrete-event simulation software. The recent integration of optimization techniques into simulation practice, specically into commercial software, has become nearly ubiquitous, as most discrete-event simulation packages now include some form of “optimization” routine. The main thesis of this article, how-ever,is that there is a disconnect between research in simulation optimization—which has addressed the stochastic nature of discrete-event simulation by concentratingon theoretical results of convergence and specialized algorithms that are mathematically elegant—and the recent software developments, which implement very general algorithms adopted from techniques in the deterministic optimization metaheuristic literature (e.g., genetic algorithms, tabu search, artificial neural networks). A tutorial exposition that summarizes the approaches found in the research literature is included, as well as a discussion contrasting these approaches with the algorithms implemented in commercial software. The article concludes with the author's speculations on promising research areas and possible future directions in practice.

Reinforcement learning of motor skills with policy gradients

“…The first simplification is the usage of the same random number generator for exploration across all seven motor primitives. This simplification is well known to introduce similar exploration over all degrees of freedom and has been shown to reduce the variance in the gradient estimate (Fu, 2002). It is necessary, as otherwise the exploration noise added in one DOF will often "fight" the exploration noise of other DOFs, resulting in very slow learning.…”

Downside Loss Aversion and Portfolio Management

“…The M-LPM n efficient frontier is the solution to expression (12) for different (a, µ). If in (12), instead of X ∈ Z we have X ∈ Z 0 , the efficient frontier is generated by only the risky assets. Different from the M-V frontier, the M-LPM n efficient frontier changes for different values of the pair (a, µ).…”

“…This corresponds to the solution to problem (12) which contains preferences exhibiting both risk and loss aversion (as in expression (6)). The optimal solution to (12) lies between the M-V and M-LPM optimal portfolios. To illustrate the computations, we use LP M 1 as the downside risk measure.…”

“…less skewed return distributions, at low return levels the optimal portfolios have distributions very close to lognormal, and therefore the difference between the M-V and M-LPM optimal portfolios are small or zero if the M-V portfolios have zero LPM. 12 Because the exponential function skews returns to the right, we make the magnitude of the largest positive jump smaller than the largest negative jump. In contrast, Liu, Longstaff and Pan (2003) used symmetric jump sizes.…”

New product development projects evaluation under time uncertainty

“…One of the great difficulties in approaching a problem through optimization techniques based on simulation (simulation based optimization) lies in the development of an efficient algorithm (Fu, 2002). As we will subsequently see, the complexity of the developed algorithm is exponential.…”