ATTENDEES

Michael Wong, Richard Dosselmann, Antony Peacock, Baran Erfani, Inbal Levi, Jesun Firos, Jolanta Opara, Kevin Deweese, Larry Lewis, Michael Chiu, Paul Preney, Rene Morell, Scott McMillan, Walter Brown

LOGISTICS

ISO meetings canceled thru Dec

CppCon - 80% online, 20% physical

C++20 – accepted in Prague; national ballots just received; waiting to work thru ISO processes

STATUS

Reinforcement Learning

LL: determining underlying technologies (e.g. tensor, neural nets)

Graph

PR: No significant change

Auto-Differentiation

(no report)

STATISTICS (P1708) Review

RD: (Overview)

RD: (Intro)

WB

    Why these functions, and not others (e.g. skewness, kertosis)?

MW

    Plan: layered delivery (simple), followed by others like kertosis

    Should they be delivered together? Separately?

RD

    Delivering similar functionality as other language (esp. Python)

    Based on existing literature

WB

    Look for existing ISO paper for statistics?

PR

     How to find them?

WB

    ISO web site. Write to them to request relevant paper(s).

MW

    I’ll take that on.

    SG6 (Numerics) may add other perspectives & requirements beyond matching other languages (hence involvement of WB)

JO

    Suggest simple function median() to call quantile with 0.5

    I think it’s counter-intuitive to have default value for quantile of 0.5

JO

    Not mentioned; general weights should be non-negative

    geometric_mean: should negative values be allowed (tricky); should be stated explicitly

    harmonic_mean: (same)

    will send you notes

JO

    quantile: the sum of weights is 1.0. Not identified in mean

JO

    Pop variance (7): exponent should be outside parenthesis

MD

    Keep stats function in <stats>

WB

    Define why you want them in their own header

   Random functionality is in <random>

    Special math functions were put <cmath>

Context: <review mean example>

WB

    Design slightly problematic

    Result=double is questionable; More effort to eval the type (e.g. from input_range R)

    Result template arg should come first (usability)

PR

    Agree with WB

    Suggest return type of auto

MC

    What’s the justification of these functions, in comparison to using lambda’s

RD

    To provide a straight-forward implementation of statistical functions, similar to other languages.

MC

    We have an opportunity to handle more specific implementations.

    The current implementations imply naïve implementations and other features may be important to be supported.

    tag_invoke [should be looked at to support different, useful implementations]

WB

    P1895

IL

    might be relevant (?)

    https://www.iso.org/obp/ui/#iso:std:iso:3534:-1:ed-2:v2:en

    (Statistics — Vocabulary and symbols)

RM

    tag_invoke.. https://wg21.link/p1895

    Also.. I wrote an isolate more recent application of tag_invoke here.. https://github.com/bfgroup/duck_invoke

Context: <var func>

JO

    NaN values: What is NaN?

WB

    quiet NaN or signalling NaN

JO

    Population vs sample: Which is more frequency used

MC

    I think you should just have one variance and pass a weight. (weight instead of choice)

Context: sorted_quantile, unsorted_quantile

JO

    sorted_ and unsorted_ are strange names:

    median is quantile, quantile is not median: suggest median (helper) function

    we always have 2 numbers; would like to have policy to choose one number

MC

    When using quantile is used for text, then it doesn’t make sense to return average

RD

    unsorted_quantile has lower restrictions on type of range

WB

    What about the concept of a sorted_input_range?

    Allow overload resolution to select the proper function

PP

    Examples might be simpler using structured bindings instead of get<n>()

MC

    Disagree with handling of text

    Nothing to do with statistics

MD

    SG19 feedback was that it was almost a side benefit.

    We should support it if it made sense and users find it useful.

JO

    I would like to see in quantile a list of quantiles I can specify

MD

    You want to specify multiple quantiles?

JO

    Yes

Context: <mode>

Context: <accumulator objects>

LL

    Is there a way for someone to do a kahan summation?

MC

    Why is mean in an accumulator a function? Why not just a member value?

WB

    That requires some thought. There are a number of values involved.

MD

    Mode returns a collection, not a single value

PP

    do we want to return a range?

    Concepts might help clean things up. Would that be helpful?

MD

   [Where are] kurtosis & skewness [used]?

MC

    They’re in Matlab

MW

    Can they be useful outside of machine learning?

    What other languages?

    Where do we stop?

MC

    Are there plans to have an implementation?

    Boost proposal?

MD

    I have a quick-and-dirty implementation

MC

    It would be helpful to propose it to boost