How to obtain weights used in policy optimization?

Use this forum to post Vensim related questions.
Post Reply
portak
Junior Member
Posts: 15
Joined: Fri Dec 07, 2018 2:43 pm
Vensim version: DSS

How to obtain weights used in policy optimization?

Post by portak »

Hello,

I have a model in which there are two annual decisions for 40 years. There is a final target at the end of the simulation and there are three criteria to consider during the simulation.

My payoff function is in the attachments. "Target reached at 2060?" is the final target.
I change the weights of three criteria manually to see the effects of weights on a performance indicator. The total of weights is equal to 1. The performance indicator is calculated based on the three criteria presented in the payoff function. These are my simulations.

I also run experiments by letting the human decision makers decide annually for 40 years. But they have the weights in their minds and not in the model.

Now, by using the Calibration tool, I would like to obtain the weights a subject gives to these three criteria. But since in my model the weights are only used in optimization and not in the model itself, I do not know how to do it. In other words, I would like to get weights used during policy optimization. Could you please help me?
Attachments
payoff.png
payoff.png (18.53 KiB) Viewed 2862 times
Administrator
Super Administrator
Posts: 4573
Joined: Wed Mar 05, 2003 3:10 am

Re: How to obtain weights used in policy optimization?

Post by Administrator »

When you run the optimization, a file will be generated with the extension .out. This can be used as a standard Vensim CIN file, so run your model using this as an input, and the resulting VDF file will contain the optimized result.
Advice to posters seeking help (it really helps us to help you)
http://www.ventanasystems.co.uk/forum/v ... f=2&t=4391

Units are important!
http://www.bbc.co.uk/news/magazine-27509559
tomfid
Administrator
Posts: 3804
Joined: Wed May 24, 2006 4:54 am

Re: How to obtain weights used in policy optimization?

Post by tomfid »

You can make the weights parameters in your calibration. If they have to sum to 1, the optimizer won't respect that, but you can normalize. Using "w" for weight and "p" for priority:

optimizer chooses p1, p2, p3

w1 = p1/(p1+p2+p3)
etc.

If people are making decisions over time (as in gaming mode), there's a conceptual problem with this: optimization implies that people know the future, which they don't in reality.
portak
Junior Member
Posts: 15
Joined: Fri Dec 07, 2018 2:43 pm
Vensim version: DSS

Re: How to obtain weights used in policy optimization?

Post by portak »

Thank you for the answers!

@tomfid: Ensuring the sum of 1 would be my next question :) Thank you for the answer.
The future in my model depends on delays and the subjects know about them.

@Administrator: I am not sure I understand how to proceed. So I have annual decisions for 40 years from the experiments. Based on these decisions, I would like to know what would be the weights used in policy optimization (not the variables) to have these decisions.
I still need to research and understand how the CIN files work but what I understood from your answer is that I can have the optimal variables, not the weights. Is that correct?
tomfid
Administrator
Posts: 3804
Joined: Wed May 24, 2006 4:54 am

Re: How to obtain weights used in policy optimization?

Post by tomfid »

It's one thing to say that participants know of the existence of delays. It's quite a different thing to suggest that they can correctly mentally simulate the outcome of those delays and any other model structure. The dynamic decision making literature shows pretty clearly that they can't, even on simple first-order tasks.
tomfid
Administrator
Posts: 3804
Joined: Wed May 24, 2006 4:54 am

Re: How to obtain weights used in policy optimization?

Post by tomfid »

Optimization is extremely useful for coming up with a benchmark policy that maximizes the objective function. However, in real game play situations, people can't plausibly optimize, because that would imply that they were fully informed about the system structure and parameters and could accurately simulate the consequences of possible decisions at every point in time.

So, what you need, in addition to the optimal benchmark, is some heuristics that make the decisions at each point in time a function only of the current system state, or more accurately, of the perceived system state. Then you can estimate the parameters of those heuristics in order to maximize the fit between model and observations.
tomfid
Administrator
Posts: 3804
Joined: Wed May 24, 2006 4:54 am

Re: How to obtain weights used in policy optimization?

Post by tomfid »

Post Reply