I would like to build a simple human-readable model based on connected concepts that can model how people - very approximately - represent and update their decision-making on a political subject as new information and experiences occur. The idea is that updating occurs due to some algorithm. I am perplexed which method to choose - or even whether this is a dead end as I increasingly fear! Anyone with expertise got any ideas?

Some models I have explored and my rough thoughts below:

Cognitive affective maps (Thagard, Homer Dixon) and they work by means of association, where each concept has an activation and valence from bad to good. If the connection between concepts is positive then the valence of the concepts becomes more similar. If it is negative, then the valences drift become opposite. Thagard was never able to come up with a convincing way for them to update (HOTCO2 is riddled with inconsistencies) and I think this is due to the model being overly-reductive. Advantage: parsimonious, captures important essence of how we think.

Causal inference maps and their ilk (Axelrod, Pearl etc). Concepts are connected in a cause-effect manner. Various intermediary nodes can be used to represent more complex constructs such as used in the argumentation literature such as AND or XOR. If concepts have a utility attached to them, then the best decision is the one that leads to the highest utility. The problem with these models is they can require too many parameters and too many nodes, leading one away from parsimony and making them a big spaghetti mess. Advantage: can represent many thought structures, intuitive if they are small.

Constraint-satisfaction models (Nerb, Kim, Glockner). Similar to a mix of cognitive affective maps and neural networks where the nodes are concepts and activation spreads across them. It's weights and values are adjusted until the network is stable and able to predict observed facts. (might be butchering this). But it fails to take into account directional inference and valence.

Bayesian nets (Pearl etc). Needs no introduction, but hard to find parameters for.

Argument mapping. Similar to inference models.

  • 1
    $\begingroup$ I think the question of "best" is necessarily opinion-based and inviting discussions, which is not suitable for the stack exchange format. $\endgroup$
    – Bryan Krause
    Jan 8, 2022 at 18:44
  • $\begingroup$ Sorry - I'm new here. So how can I get an answer as to which method is best and whether there are alternatives to those I've described. If not stackexchange then where? $\endgroup$ Jan 8, 2022 at 18:50
  • $\begingroup$ You could read papers where people have tried to do something similar to what you want to do. No model is perfect or correct; some are useful. $\endgroup$
    – Bryan Krause
    Jan 8, 2022 at 18:54
  • $\begingroup$ I've read papers and books on this subject, built the models myself and so on. Perhaps I'm in the wrong place? $\endgroup$ Jan 8, 2022 at 19:01


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.