THE DNA OF BETTER DECISIONS
Through the use of massively scaled evolutionary computation (EC), we have created an intelligent system we call Evolutionary Intelligence™.
Our evolutionary intelligence technology is breaking new ground in the modern AI era where we are applying previously academic work to create breakthroughs in the fields of investment and medical research.
Inspired by biological evolution, Evolutionary Intelligence generates specialized programs known as agents that affect solutions to proposed problems. Because EC is “embarrassingly parallelizable,” it can work semi-independently on individual CPU cores, loosely coordinated through asynchronous communications between many nodes.
In the past, EC typically converged rapidly, within a few hours on a few CPUs. Using our patented technology, we have managed to scale EC to run through trillions of generations across millions of CPUs to create a system powerful enough to solve some of the world’s most complex problems.
But how does our massively distributed EC system work?
Our AI system begins by generating—and then comparing—a diversity of candidate agents (genes) to distinguish which ones are better suited to solve a particular problem. A “fitness score” is assigned to each candidate based on how well it performs compared to its peers. Note that this is a relative measure, making it unnecessary to know the fitness score of the best possible candidate.
The first population of candidate agents all likely perform poorly as they are generated randomly (step 1). But as the system evaluates them against specific training data, some prove less bad than others (step 2), so the system keeps these and eliminates the remaining candidate pool. It then makes use of components of the better candidates in order to generate a new population (step 3). This process is repeated millions of times. Through masses of generations, massively distributed over millions of CPUs, the system gradually converges on solutions, resulting in code that works to solve the complex problems we see in trading and healthcare research today.
Sentient’s EC allows the fitness score to comprise more than just a single measure (objective), as not all systems are optimized on a single axis. In trading, for example, we are interested in returns, as well as the dispersion of the returns. The system can also measure and reward candidates on diversity or novelty of their behavior.
Sentient’s EC is also geared for tackling problems in which the fitness score is not easy to compute, such as data problems in which the only way to assess a candidate is to apply it to as many data samples as possible, aggregating the results. In Sentient’s EC, this assessment is made incrementally, in a highly distributed fashion, with more promising candidates being validated on more data samples, increasing the reliability of the solutions.
BABAK HODJAT, CO-FOUNDER AND CHIEF SCIENTIST