paulws / Banish

Banish
1 stars 2 forks source link

Setting probabilities in a node #4

Open ramstat opened 10 years ago

ramstat commented 10 years ago

These are my first thoughts about the data entry side of node probabilities (not the elicitation of these values).

It is sensible to flag up when the probability distribution is not valid, i.e. the individual probabilities do not sum to one. The options appear to be to get the user to request validation of the distribution or for it to be automatically validated each time a cell is updated. If might be worth stopping the node being updated with an invalid distribution?

When we have a node with two states then once one probability is specified the other is automatically defined. However if we have three (or more) states then the BN software could provide:

1) The option to complete the last cell as one minus the sum of the other probabilities. One danger with this approach is that the user could be tempted just to lump the remaining probabilities into the last category. 2) Scale the probabilities so that they sum to one. If the user has specified the probabilities for each category relative to one another this might be ok. 3) No short cuts so the user has to explicitly make the distribution valid.

Conditional probabilities: so if we have two nodes where the probability distribution in one is conditional on the states of another node the user may end up in situations where a new state is added. The other node will then need to be updated with the new conditional distribution. What is a good way to flag this up and to get the BN resolved before running calculations?

Hope what I have written makes sense!

paulws commented 10 years ago

Thanks for your input. We will improve the probability input along the lines you suggest. We also have to provide the facility to input the data using a probability yard stick - which will open up more issues with entering inconsistent values etc. So we could have user preferences e.g. entering probabilities manually, no short cuts, using the yard stick etc.

The second bit you raise - at the moment the colour coding of the boxes indicate when a new distribution is required. This could be made a bit more obvious. I'll try and come up with an improved display.

Paul