MDL can be used to create a model (a compressed but representative version of the real world) without truth as a concept. This model can be used to predict the future. Why is a prediction based on a model without truth better?
In MDL models are created based on ‘historic’ data. The more data the better the model will fit the problem underlying the data. (The individual parts of the data must have some relation to each other, otherwise we would be creating a model for randomness and any prediction will fit the model.) With this model we can predict if certain instances (called hypothesis) of our problem fit this model. This is not a matter of finding truth, but of finding an acceptable probability it can be considered part of the model.
If truth would have been involved we would have been trying to prove the hypothesis to be 100% in accordance with history. This does not result in a a high rate of predictability. It also does help gain insights in the relationship between the hypothesis and model. This does not help in qualitatively investigating the hypothesis.
So using truth in modeling scenarios is like, to paraphrase Arjan, ‘walking backwards into the future‘. And this is exactly what we not want. We want to ‘imagine’ plausible scenarios that inspire. For that we need tools to help build scenarios and tools that help compare scenarios to a modeled world and tools that help compare scenarios to other scenarios in such a model of the world.
For these reasons the Minimum Description Length principle is extremely interesting. I think this could be an inspiration to our development of methodologies and tools. But it could very well be the foundation for a profoundly new scenario building ‘ecosystem‘.
Image
I need a little bit of help in understanding this correctly
We have a certain problem, let’s say for instance the end of oil (correct me if this is not a good example). In MDL, a model is then created based on historic data. This model ‘describes’ the behaviour of a system (a.k.a. society?). Now, we come up with a certain hypothesis, say: the end of oil will lead to the invention of new means to fuel our cars. We put this in the model, and we get back a probability of this actually occuring… Is that the way it works?
Secondly, why do you see this as so much better than morphology? And isn’t ‘walking backwards into the future’ exactly what happens if you base ‘everything’ on historic data?
yes, your way of describing the process is like i see it. it is not so much about the probablity, it is about if it is part of the model or not. even when it is a strange hypothesis, say “oil will become abundant because we are going to farm it in greenland with new technology” might be part of the model.
this is different from morphology in key area. in morphology you extract possible hypothesis out of historic data. there is now way to ‘invent’ a hypothesis and reason about it with help of morphology. in mdl you can come with some hypothesis and ‘map’ it to the model. this way modeling the future is about creativity and prediction. where morphology is about history and elimination.
i think mdl is not walking backwards into the future. in mdl you construct a model from historic data (research, understanding.) than you ‘invent’ a hypothesis that you check to the model. this is much more like our freedomlab way of working (the whale.)
I wonder what John’s view about this is…
To add to Jorgen’s comment. The problem he states is a typical form of financial modelling on which company strategies are based and new products will be introduced. HIstorical data is always the starting base for any kind of forecasting, but wil never be the end point of decision making. Based on historical data future prices will be defined - in the case of depletion of oil very high prices. However, the historicla trend will ease out in to a more or less straight line (http://www.globalsecurity.org/military/intro/images/oil-1997.gif). Based on the indicative price, strategist will design a financial model (e.g. oil reserve as the decision variable) and run 1000 scenarios with special software. The outcome of this scenrio analysis will be the input for a brainstorm session with people of different background and experience (e.g. Head of exploration and Head of Natural Sience). The discussion will have to be around the parameters (How much will the oil reserve deplete over the years). By changing the parameters new scenarios will be defined. After choosing a few the scenarios will be compared to the introduction of new products like fuel cells (to power cars). Simply said if the new product will create a higher profit than the old product (oil), the development will start and an infrastucture for fuel cells will be constructed and a fuel-cell powered vehicle will become profitable. Hope this helps.