Evolutionary algorithms have revolutionized the way that we optimize devices in RF and nanophotonic engineering by injecting population dynamics and selective pressure into optimization procedures. The Covariance Matrix Adaptation Evolutionary Strategy (CMAES) is a preeminent example of such optimization algorithms due to its extremely attractive convergence characteristics and its robustness. CMAES works by iteratively updating a population of device designs, selecting a subset of best solutions each time. What makes CMAES unique is its use of a sampling distribution (multivariate Gaussian) to select the next generation of designs. Its carefully derived update procedure for this distribution’s parameters (mean and covariance) allow for the entire set of parent designs to contribute to the next generation at the same time.

For most engineering problems, there is rarely one “best” solution. Competing goals mean that there are often multiple designs which trade-off advantages and disadvantages with one another. Multi-objective Optimization (MOO) is an optimization strategy which directly accommodates this reality, allowing for optimization when multiple, potentially competing cost functions. Rather than forming a single “figure-of-merit” which weights costs from different objectives, MOO treats them holistically, employing unique sorting strategy called non-dominated sorting to find the “best tradeoff” (called the Pareto set) from a given set of candidate designs.

Inspired by the real-world behavior of ants, CEARL researchers developed the Multi-Objective Lazy Ant Colony (MOLACO) algorithm to optimize freeform 3D meandered metamaterial structures. MOLACO is an important extension of classical ant colony optimizers and improves upon them by introducing the concept of "lazy ants", a real phenomena that colonies employ to exchange tired ants with ones in reserve. This manifests itself in the algorithm as a novel way to achieve a variety of different meander lengths. Furthermore, by extending the algorithm to support multiple simultaneous objectives, users are able to optimize metamaterial structures for competing objectives such as SWaP (size, weight, and power) and optical performance.

Real engineering problems not only have constraints which limit the range of acceptable design values, but also tolerances which limit the granularity these values can achieve. Without the inclusion of fabrication tolerances (uncertainties) optimizers may find potentially amazing solutions that cannot be realized in practice due the sensitivity of the design falling outside achievable tolerances. Therefore, an optimizer than can provide designers with a set of solutions that map out the tradeoffs between achievable performance and design robustness (or requisite tolerances needed) is potentially paradigm-altering. To this end, CEARL researchers have pioneered the use of surrogate models and deep learning techniques to enable a new a process called "robustness optimization" where designs are explicitly optimized for insensitivity to fabrication uncertainties. The first implementation of this was developed in our Multi-Objective optimization w/ TOLerance (MOTOL) algorithm.

The Window Driven Optimization (WDO) Algorithm is a stochastic population-based global optimizer inspired by atmospheric motion. The Wind Driven Optimization (WDO) technique is a population based iterative heuristic global optimization algorithm for multi-dimensional and multi-modal problems with the ability to implement constraints on the search domain. At its core, a population of infinitesimally small air parcels navigates over an N-dimensional search space following Newton's second law of motion, which is also used to describe the motion of air parcels within the earth's atmosphere. Compared to similar particle based algorithms, WDO employs additional terms in the velocity update equation (e.g., gravitation and Coriolis forces), providing robustness and extra degrees of freedom to fine tune the optimization. 

Swarm intelligence, one of the newest optimization techniques, was first introduced by James Kennedy and Russell C. Eberhart in 1995. Kennedy and Eberhart studied the behavior of individuals from the social point of view: individual behavior in groups. PSO simply mimics the success in optimization as a group. PSO is a stochastic, population based, robust multi-objective search algorithm. It does not require prior knowledge of the shape of the search space. The algorithm is very simple and easy to implement and PSO has recently been shown to be useful for solving complex electromagnetics and antenna optimization problems.