I have a short thought, stemming from a combination of projects that I’m working on at the moment, and I want to share it.
The current trend towards Causality in AI is very attractive to people like me. It matches our personal biases and views of the world. However, it is lacking a natural heuristic. How do we decide how much resources to devote to alternative models of the world, as we gather evidence as to their accuracy?
Like I say, I have a number of parallel projects, many of which address exactly this question on technical and biological levels.
There is something from the world of business, studying entrepreneurship, which might be a better heuristic than any normative model I can come up with. Effectual entrepreneurship is a perspective on entrepreneurship, studying highly successful repeat entrepreneurs (eg. Elon Musk), which establishes control, rather than planning, at the core of entrepreneurial activities.
Basically, planning is appropriate under stable conditions which resemble management in a large company. Entrepreneurs focus on the means that they can control and then observe where they can steer the project. They then decide, based on other factors, whether to continue or not.
This resembles the scientific method.
I’m not kidding. If we look at the history of development of scientific thought it is becoming clear that the amazing breakthrough in science is actually the acceptance that we must prove our theories. Every scientific theory of the world is essentially a circular model which is then short-cutted by an experiment which takes a slice of the real-world and sees whether the experimental result matches the theory. But I don’t really want to discuss scientific theory here.
Instead, let’s think about Causal models. The picture that I and people like me tend to form when thinking of causal models, for AI, is essentially an advanced critic in an Actor-Critic model. The critic maintains evidence in support of the current world-model, and it also devotes some resources towards developing alternative world-hypotheses. If one of these alternative world-models has greater supporting evidence than the existing critic, then it replaces it for the current task.
The hard thing is we have to come with some kind of external hypothesis as to how many resources it is worth devoting to these alternative world-views. There are an infinite number of hypotheses about the world. It is not worth gathering evidence for all of them. There does not seem to be an encoding which will appear which will ever overcome this infinite issue. So we resort to either resource heuristics, or biological inspiration. I currently have a project in each of these directions.
The inspiration from effectuation leads me to think that there may be a similar short-cut to that discovered in the invention of the scientific method. Maybe Causal Models should not focus so much on all models (ok, it’s clear they shouldn’t) but rather on what the agent can control!
This is essentially what humans do anyway. It might also be an explanation, at some stage, for why many people in society disengage mentally from things they cannot control. It’s a sensible use of resources.
So the heuristic becomes, things which I can impact (control) get space in my critic-space. This might be in direct proportion to how much I can impact upon them. And, everything else is ignored.
This whole concept is essentially a reformulation of the results of my preprint from a number of years ago. What is different is the focus on using degree of control as the deciding factor rather than variance in performance.