During my years as a quant trader, I found myself trying to automate my own job.
As it turns out, the market is a formidable adaptive adversary and profitable trading models tend to expire because relationships between assets are always in constant flux. Whenever you derive a predictive model for a given asset, it is basically impossible to tell if the model will remain predictive for one day, one month or one year. If you stop to think about it, it makes sense because we are always modeling the past of a dynamic non-stationary system.
My goal was to create “wide-AI” (not narrow but not strong either) within the limited domain of trading, the idea was to build a system that was not only capable of trading any asset but that was also able to decide by itself which assets to trade and which data streams to use without any human intervention.
My job as a quant was to use several statistical and machine learning tools to derive trading models. The question I wanted an answer for was: how could I build a computer program capable of replacing myself?
To accomplish this goal I built a multi-layer system where the top layers would analyze the data and generate a population of independent programs that would in turn attempt to maximize the value of a utility function. The programs that composed the populations in the bottom layers were themselves recombinant and made of several different machine learning “blocks” and data transformation pipelines.
I was eventually successful and the most important lesson I learned is that when you create a system capable of dynamically integrating different adaptive tools at runtime, you may end up with something far greater than the sum of its parts.
While remarkable, the system I built was still confined to a single domain. After reflecting over the outcome of my research, I realized that within the limited scope of the financial markets, I had built something I decided to call a “domain survivalist”.
Considering the tradable market to be its environment and the joint set of available inputs and outputs to be its embodiment, I had created an agent capable of bootstrapping itself to its “body” and “environment” in order to survive and maximize the value of an arbitrary utility function.
The same system could just as easily trade Oil, Corn, Euros or Google stock. However, even though it was a bit more flexible than what I could have achieved with more traditional methods, its operational range was still vexingly narrow – trading was all it could ever do.
Towards the creation of a Universal Survivalist
Now let’s take the concept of a survivalist AI to the next level. If we set our hearts and minds to it, could we build such thing as a “universal survivalist”? Could we build an AI agent capable of bootstrapping itself to any arbitrary embodiment and find ways to exploit its environment in order to maximize the value a computable utility function?
If the body of the agent is defined as the set of inputs and outputs that are available as means to respectively sample and actuate over its environment, the concept of “embodiment” becomes rather flexible and can extend itself nicely to include AI agents without any sort of physical components.
While creation of such universal survivalist system would not give us HAL 9000 (because language is too complicated), it could certainly give us the algorithmic underpinnings for AI that is versatile enough to bootstrap itself and intelligently control any arbitrarily complex system, be it a plane, a car or a firewall.
The same underlying architecture could then be used to infuse different machines with varying degrees of useable intelligence. A survivalist “born and raised” inside a car could be trained to obey traffic laws the same way police dogs are trained to serve and protect. Similarly, another identical survivalist instance that was instead attached to a security system could be trained to protect a particular location, such as a bank or a hospital. Once trained, any survivalist could then be cloned into as many similar “bodies” as needed.
Indeed something to think about…