Equation Node Documentation?

[SteveM]

Genie Genii,

I'm running version 2.0.3259.0 I'm only a casual user right now. I noticed the new Equation Node type but can't find any documentation for it's use. It appears that the Help files have not been updated.

Can you point me to something?

Thanks,

SteveM

P.S. Go Steelers!

[marek [BayesFusion]]

I'm running version 2.0.3259.0 I'm only a casual user right now. I noticed the new Equation Node type but can't find any documentation for it's use. It appears that the Help files have not been updated.

Can you point me to something?

Steve,

I'm afraid we have no documentation on that -- not enough manpower to work on the Wiki at the moment . They are really pretty self-explanatory but I will be glad to try to help if you have any specific question.
Cheers,

Marek

[Per Närman]

Hello!

Using the Equation node sounds promising in many ways. Do you have any idea when its documentation will be released?

Kind regards

Per

[shooltz[BayesFusion]]

Do you have any idea when its documentation will be released?

No idea, but I'll be glad to answer any questions here.

[Per Närman]

Excellent, thanks.

A very fundamental question to get things started: what is the underlying mathematics behind the equation nodes? Are they somehow based on Bayesian networks, or are they using regular algebra?

A second question: When you update your probability distributions and have a look at the value of the nodes it looks like a Monte Carlo simulation. Is this a correct interpretation?

Thanks!

/Per

[shooltz[BayesFusion]]

If the model is fully deterministic we're using algebra. There's lot of work ahead here - we're not even close to other packages when it comes to symbolically solving the equations.

When some of the equations contain probability distributions, we're switching to stochastic algorithm. Without evidence we run forward sampling, with evidence it's more sophisticated sampling algorithm. In such case the results are stochastic samples. There's also an option of transforming nondeterministic, equation-based model into plain BN by discretizing it. GeNIe can do it internally; it'll create the temporary BN, run the inference and retain the results for each node (the probabilities of each discretized state). Users can also create discretized BN from equation-based models manually for explanatory/debug purposes.

[Per Närman]

Thank you for your answer! I have even more questions...

1) As for the mathematics: if I understand it correctly the equation nodes are not using continuous Bayesian networks, rather they are based on whatever mathematics fit the different situations at hand (algebra sometimes, simulations at other times etc.). Am I right?

2) I am looking for a way to use MIN and MAX functions in for continuous variables. Do you think it is possible to specify those functions using the "Conditional Functions"? And if so, is there an example of the syntax for writing the "If" and "Switch" statements?

3) What are the "Switch" functions anyway, and what is their intended use?

Regards and thank you!

Per

[shooltz[BayesFusion]]

I'm forwarding question 1 to other members of the team.

2) I am looking for a way to use MIN and MAX functions in for continuous variables. Do you think it is possible to specify those functions using the "Conditional Functions"? And if so, is there an example of the syntax for writing the "If" and "Switch" statements?

Current GeNIe build already contains Min/Max functions, so you don't really need to bother with if/switch. Anyway, here are the examples:

If(cond, tval, fval) evaluates the 'cond' argument. If evaluation returns nonzero, the return value is 'tval'; otherwise 'fval' is returned. The value of the expression below is 7:

Code: Select all

If(2-2, 5, 7)

Switch(idx, val1, val2, ...) evaluates the 'idx' argument and uses it as 1-based index to select valN as its return value. The value of expression below is 333:

Code: Select all

Switch(3, 1, 22, 333, 4444, 55555)

[shooltz[BayesFusion]]

2) I am looking for a way to use MIN and MAX functions in for continuous variables.

BTW, you can also use the 'Trim' function if you want to limit the value of expression to specified range.

[marek [BayesFusion]]

1) As for the mathematics: if I understand it correctly the equation nodes are not using continuous Bayesian networks, rather they are based on whatever mathematics fit the different situations at hand (algebra sometimes, simulations at other times etc.). Am I right?

I believe that it is wrong. There is a whole spectrum of Bayesian methods and their distinguishing factors is that they are capable of (or they just do) updating of probability distributions in light of evidence. GeNIe equation nodes can do that. Please note that you can specify noise terms within any equation. Discrete and continuous distributions/equations are, in fact, a continuum that, regretfully, few people in AI perceive as such. Does this clarify your question?
Cheers,

Marek

[Per Närman]

Thank you again for your response.

We are using BNs and GeNIe in our research that concerns evaluating IT system properties. The equation nodes increased ability to reason using (what I sloppily refer to as) continuous variables makes it a whole lot easier to compute performance metrics for instance. I have understood that there is a plethora of written material on BNs but the equation nodes seem to be a bit different from what you usually read about in the BN world, aren't they?

We are very interested in using the equation nodes in the work we do, but it would be great if we could describe what the equation nodes actually do, mathematically speaking. So I think what I really want to know is if there is some paper or other source which describes the algorithms that are used?



Kind regards

Per

[marek [BayesFusion]]

Per,

Sorry for missing this -- I've started to clean my mailbox and noticed that I have never responded to your query. The timing makes this answer probably irrelevant but let me try anyway. Continuous nodes are not any different than discrete nodes except for the domain of variables. Discrete nodes have very limited domains (e.g., true/false, high/medium/low, etc.). Continuous nodes can assume any numerical values (their domains are always numerical). What a Bayesian network does is computing the probability distribution over the domain of a variable and it does not matter whether the variable is discrete or continuous. Does this help in reconciling the difference in your mind?

Having said that, I would like to point out that GeNIe does not allow you at the moment to mix these two types . In theory, this restriction is not needed. I guess, it is all in early stages of development and we don't have algorithms handling both at the same time -- this is the reason. Please look, however, at discretization of continuous nodes -- this lets you do inference under most circumstances.

I hope this all helps.
Cheers,

Marek

[gingerauto]

Hi everybody;

I'm a beginner in the field...i have to model a Dynamic Bayesian Network..There is no documentation about this, except one incomplete example...
Can i find documents Somewhere else?

I have other question , how can we add an arc from equation nodes to others!! it is possible? Otherwise, how can we model the dynamic evolution of variables(node)?

Thanks in advance

[gingerauto]

HI,
I want to know other things!!
When we want to model an equation node that changes in the time, how can we introduce time variable in the node equation!!

[marek [BayesFusion]]

When we want to model an equation node that changes in the time, how can we introduce time variable in the node equation!!

I'm afraid we don't support that (yet ). You could, of course, have an explicit variable "time" in your model and use this variable in equations. That variable would be the parent of variables that use it. I realize that this may be somewhat cumbersome but I can't help more at this point.
Good luck!

Marek