How can we design experiments in quantum physics? Generally, the strategy involves using the vast knowledge amassed by quantum physics researchers and combining this with our intuitions and our creativity. This method has been fantastically successful over the years and has given us experiments that help unlock both the technological and the foundational magic of quantum physics. But it is not always easy. Quantum physics is notoriously counter-intuitive, so sometimes our intuitions are not a good guide for how best to design an experiment. And sometimes past experience is not always the best guide to creating novel new designs, particularly when our experiments intend to probe a new and untested area. But how else might we design quantum experiments? In a previous post we introduced a method we are developing that utilises techniques from AI, machine learning and meta-heuristics to design quantum experiments. In effect, we are letting computers design experiments for us.
To be more specific, our task is the following:
Given a set of quantum-optics experimental equipment, what is the best way of arranging the apparatus to create a quantum state with certain desirable properties?
What we mean by “desirable” depends on the application in mind. We are interested in designing quantum states of light, which can be useful range of applications such as quantum computing, making high precision measurements, and quantum cryptography. Now, the apparatus we are using can be broken into three categories: states, operations, and measurements. Roughly speaking, we take some states, act on these states with some operations, which modify the states and cause them to interact with one another, then we perform some measurement. This is getting rather complicated, so to simplify things we will use Lego!
The meanings of all the symbols in the white boxes is not so important for the purpose of this blog, but for those already familiar with quantum optics we provide a glossary at the end of this blog that says what all these different symbols means. If we now think of this in terms of Lego, our question becomes:
Given a collection of Lego pieces, what is the best way to arrange them to produce a construction with certain properties?
Just like in quantum optics, there can be a variety of desired Lego constructions, such as making something strong, something beautiful, or making something that looks like a real-life object such as a police car. And just like with quantum experiments, the usual way to design a Lego construction is to use creativity, prior knowledge, and intuition.
But how could a computer design a new Lego construction? The technique we use is known as a genetic algorithm. Genetic algorithms are designed to mimic natural selection, and they work as follows:
First, the computer makes a collection of completely random constructions. This collection is known as the initial population:
We then assess the various designs, giving them all a score:
Next, we select the best designs, and throw away the ones that aren’t so good:
This leaves us with a smaller subset of the population, which we call the parents:
We will use these parents to create a new population, known as the children. First, we take the very best parents, and copy them without any modification. This produces the elite children:
Second, we breed some of the parents together to produce crossover children:
Thirdly, we make small changes to some of the parents, producing the mutation children:
This leaves us with a new population, which hopefully should be significantly better than the initial population:
We then repeat this process:
Each cycle is known as a generation, and after repeated generations we create better and better individuals. Just like in nature, eventually we end up with individuals that are highly suited for our goals (try and work out what the goal was here!):
And the result is the same for us: after a number of generations, our genetic algorithm produces a range of new quantum experiments that often outperform the human-designed experiments!
Acknowledgement: It should be stressed that this post is based on a poster by Rosanna Nichols, who should take most of the credit! A paper based on the genetic algorithm introduced here will be available shortly, but the earlier paper can be found here.