4 ms·

Genuinely curious: what's a realistic use case for something like this?

by qeternity 10y ago

Abishek_Muthian 10y ago
NLP,AI etc.
- swsieber 10y ago
  I think Family history fits in here too (seriously).
  - tr352 10y ago
    Interesting. Could you elaborate on what you have in mind?
    - swsieber 10y ago
      I'm think of automatic grouping of data points into people profile. For instance, when searching for John Doe across all census records, provide profiles of people, backed up by multiple sources. This would be achieved by seeing if a group of two references are likely to be referring to the second person , and of they are, merge the groups.
pizza 10y ago
Stochastic heuristics/sanity-checks
tr352 10y ago
RankPL can be used instead of probabilistic programming for problems where probabilities are unknown, but where it's still possible to distinguish normal from surprising behavior of a model. Admittedly, this is still quite abstract. I'm working on more concrete use cases.
- santaragolabs 10y ago
  Here's one that is maybe more concrete. And I hope I'm understanding everything correctly. Say you're a startup running your infrastructure in AWS. You spread it out over three different regions and within each region you use 2 availability zones. Your network load is automatically balanced over these three geographical regions. Now an earthquake happens in one region and although it's unlikely both of availability zones within that region go off line (fiber to the region is cut, power-loss, whatever). This means the entire region goes offline. If modeled properly you should now be able to figure out what the consequences of this will be for the entire infrastructure. Will you be able to stay online if surprising behavior (an entire region going offline) happens? Of course the big issue here is always mapping real world scenario's onto models that fit well enough. EDIT: It's a matter of taking the "nasty integral" part out of it as per nerdponx in another comment on this thread. This can really help with doing Fault Tree Analysis for example as the statistics solving part there has always been a big problem for systems big enough (MCMC solvers help only to a degree).

Genuinely curious: what's a realistic use case for something like this?

Abishek_Muthian 10y ago

NLP,AI etc.

swsieber 10y ago
I think Family history fits in here too (seriously).
- tr352 10y ago
  Interesting. Could you elaborate on what you have in mind?
  - swsieber 10y ago
    I'm think of automatic grouping of data points into people profile. For instance, when searching for John Doe across all census records, provide profiles of people, backed up by multiple sources. This would be achieved by seeing if a group of two references are likely to be referring to the second person , and of they are, merge the groups.

pizza 10y ago

Stochastic heuristics/sanity-checks

tr352 10y ago

RankPL can be used instead of probabilistic programming for problems where probabilities are unknown, but where it's still possible to distinguish normal from surprising behavior of a model. Admittedly, this is still quite abstract. I'm working on more concrete use cases.

santaragolabs 10y ago
Here's one that is maybe more concrete. And I hope I'm understanding everything correctly. Say you're a startup running your infrastructure in AWS. You spread it out over three different regions and within each region you use 2 availability zones. Your network load is automatically balanced over these three geographical regions. Now an earthquake happens in one region and although it's unlikely both of availability zones within that region go off line (fiber to the region is cut, power-loss, whatever). This means the entire region goes offline. If modeled properly you should now be able to figure out what the consequences of this will be for the entire infrastructure. Will you be able to stay online if surprising behavior (an entire region going offline) happens? Of course the big issue here is always mapping real world scenario's onto models that fit well enough. EDIT: It's a matter of taking the "nasty integral" part out of it as per nerdponx in another comment on this thread. This can really help with doing Fault Tree Analysis for example as the statistics solving part there has always been a big problem for systems big enough (MCMC solvers help only to a degree).