Intuitive noisy averager

The front end.
Post Reply
charlie
Posts: 66
Joined: Wed Aug 09, 2017 10:55 pm

Intuitive noisy averager

Post by charlie »

I guess many have used this but it's no harm to put it here again. Often in knowledge elicitation we need to factorize a complex problem or system down to the level we feel comfortable to assign a probability. For this purpose I found I need a highly intuitive noisy averager that intuitively links a parent variable to identified parent variables with simple weighted averaging (could be deterministic or noisy). Also, expert knowledge usually focuses on variables that experts are able to describe (qualitatively or quantitatively) in association with experiences or proven theories / reasonable assumptions. Leaks are not what they felt comfortable to specify - they either choose to ignore or add new variables till the exhaustion of their knowledge.

Under this circumstance, I used the attached arrangement of the very powerful noisy adder:
* Include a Not for Use (NfU) state in the parent and child nodes
* Set the weight of Leak extremely small (insignificant).

This weighted average-based factorization can go on and on but usually 4 or 5 layers are more than enough as I found in my experience. I guess this may not be liked by GeNIe creators but it really helps make things easier to elicit and communicate.

Charlie
Attachments
IntuitiveAverager.xdsl
(3.39 KiB) Downloaded 323 times
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Intuitive noisy averager

Post by marek [BayesFusion] »

Hi Charlie,

Thanks for sharing this with the community. I have not used the Noisy-Adder/Average model in practice, so I am really pleased to see that it has useful applications. The attached model seems intuitive. I believe there is no difference between the two disconnected networks, correct? The NfU states in the parents have zero probability, so they will always stay zero. I understand why you have them -- you need a distinguished state in the parent nodes, i.e., the state that has no influence on the child (Noisy-Adder). Should they, however, have some non-zero probability? Have you applied these ideas in a non-abstract situation? Then the NfU states may have some real-world meaning. Good work anyway!

Marek
charlie
Posts: 66
Joined: Wed Aug 09, 2017 10:55 pm

Re: Intuitive noisy averager

Post by charlie »

Hi Marek

Thanks for responding. Yes, the two nets are the same.

I found noisy-adder is extremely useful in eliciting knowledge from a non-pure-technological/physical field e.g. renewable energy development where surely are technological elements and a lot of other physical presences but also, at least equally importantly, human and social factors such as people's motivation of being green and the influence of social, financial and marketplace preferences. Many causal relationships in the game-play can neither be satisfactorily described as noisy-max or noisy-min or even other types of gates that can be relatively simply constructed from them. On the other hand, a full conditional probability table is way too detailed for the existing amount of knowledge to handle. A noisy-adder/averager is thus useful.

I attached a hypothetical case of a business' motivation to become more sustainable as an example. It's quite comfortable for a business manager to identify or affirm the presence of influences from three key decision factors - net profit, biz growth opportunity, and social pressure. Upon elicitation, he further identified the social pressure is from peer pressure and customer requirement. It is also quite intuitive for him to weigh the influences across the parent variables. But he and maybe the knowledge elicitator usually won't be able, or won't feel comfortable, to identify the exact logic type (max or min or another, which might indeed be very complex in nature at this level of knowledge).

Hopefully the above makes a sense.

Cheers
Charlie
Attachments
GreenBiz_NoisyAdder.xdsl
(3.83 KiB) Downloaded 378 times
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Intuitive noisy averager

Post by marek [BayesFusion] »

Cool, thanks for sharing this model with other GeNIe users. One suggestion, which is general, not necessarily about your model, as you may have other motivation for this. In the nodes "Net profit" and "Long-term biz opportunity" you have states that have zero a-priori probability. When they have zero probability and are, thus, impossible, you can just as well skip them -- they will never achieve non-zero probability (in Bayes theorem, a zero can never become anything else, as to calculate the posterior, you always multiply by the prior and zero will always be zero). Otherwise, an interesting example.
Cheers,

Marek
charlie
Posts: 66
Joined: Wed Aug 09, 2017 10:55 pm

Re: Intuitive noisy averager

Post by charlie »

You're right, Marek. It's my oversight not to assign an appropriate non-zero probabilities to the "low" and "none" states for the two nodes. There was also an error in Social Pressure where I accidentally assign a 1 to NfU state. I re-attached the net with these corrected.
Charlie
Attachments
GreenBiz_NoisyAdder (4).xdsl
(3.83 KiB) Downloaded 311 times
Post Reply