2. Where to start?


Where to even start such an ambitious project? The task is daunting, but we can at least begin to think about it and share ideas on how it could be possible. It might take ten or it might take fifty years before we will get useful results, but the longer we're not even contemplating it, the longer we're postponing it.

It seems natural to start small and simplified, and then increase the size and complexity step by step. The calibration of new elements and parameters will have to be automated in a machine learning way.


First steps

Program a simulated character with a limited amount of time to sleep, work, eat and enjoy leasure time. Spending time at work reduces the time it can spend on hobbies, but gives it the money to do more fun stuff and eat more delicious food. Too little sleep decreases the value of the time it's awake. If we translate everything it does and consumes into a common unit of value (a utility function), optimization of its quality of life is just a max-min problem.

Create another character with slightly different preferences and let it compete with the first one for some limited resources. Shortages will increase the cost of that commodity. Now we have two max-min problems that will reach an equlibrium with each other. Keep adding characters.

Early on, a small number of agents will interact and only have a limited number of choices they can make. Adding degrees of freedom will increase the model's similarity to the real world. (In the beginning, institutions and companies will also have agency, but when the simulation grows, this can be replaced with the agency of the people deciding over those organisations.)

I hope these early models will be open source so anyone can download them and tinker with them, to get a feeling for the possibilities and limitations of the simulations. But to really make progress, experts from different fields need to colaborate. Economists have experience in modelling human behaviour, mathematicians can handle the statistical calculations, psychologists can help with characterizing individuals, machine learning experts and game developers will program the simulation.

The hard part is to train the model by comparing its result to real life data. I will discuss this more in future posts.

When the computer model reaches a certain complexity, it will no longer be possible to understand how it calculates the results or to improve it manually. It will have to be self-improving and we will see it as a black box, providing predicitions with increasing accuracy, as it re-evaluates itself against what happens in the world.


Relevant Wikipedia articles:

Agent-based model

Multi-agent system

von Neumann-Morgenstern utility theorem



Comments

  1. "Simulating a better future"Thursday, September 03, 2020 11:37:00 AM

    I found this fitting quote by Anders Sandberg:
    Pipe dream technology is about mapping the space of desirable but as yet impractical technologies, learning what makes them impractical or hard. That makes us ready for when solutions to those factors arrive, and may focus attention to figuring out the factors.

    ReplyDelete

Post a Comment

Popular posts from this blog

  3. How do you describe a person? We need to be able to calculate an agent's behaviour from a limited set of quantifiable parameters. So how do you define a character's personality? It's the fact that we are different, that creates the problems we have in society, but it also creates positive dynamic forces and makes life interesting. This is what the field of psychology is all about, and I must say I think they have failed to accomplish the fundamental task of figuring out how we differ from each other and how much. A popular but rather meaningless pursuit, is to divide people into types. That's a too simplistic approach. Let me try to identify the many different layers of our personality that affects how we live. First we have physiological needs – for food, sleep, warmth – and they are basically the same for everyone, so they are not good measurements to define an individual. At the next level we have psychological needs, which differ a little more from person ...
Read more
  5. Building the framework for a simulation. Before we start to populate our simulation with agents, we should understand how the world works. Economists have produced innumerable theories about the laws that govern societies – often idealizing and simplifying the world to be able to express them in simple equations. I would like to see another approach, using machine learning to deduce how the economy functions. Take a country with good statistics and describe every county or municipality (except a few that are used to check the models' predictions). Begin with basic parameters like demographics (number of inhabitants, age, born and dead last year, age of death, moving in and out, occupations, distribution of income, unemployment) and production (agriculture, natural resources, industry). From there it's easy to find direct correlations between parameters (we don't need to worry about causation). There are also methods to find complex correlations and patterns. When...
Read more