Designing and Creating applications built on R Richard Pugh, Andy Nicholls & Chris Campbell 23rd October 2012

Thank you for the invitation to speak tonight

Andy Nicholls Senior R Consultant

Chris Campbell Senior R Consultant

Richard Pugh Principal R Consultant

& Co-Founder

Agenda

•  Who are Mango Solutions? •  Why Build Analytic Applications on R? •  Formal R Application Development •  Case Studies •  The R Community •  Discussion

Who are Mango Solutions?

Overview of Mango Solutions

•  Private Company formed in 2002

•  Global Team of ~70

•  Cross-Sector Software and Services

•  ISO 9001 Accredited

Located here ...

Bath, UK London, UK Shanghai, CN Basel, CH

Spend a lot of time here ...

The Beginning: October 2002

•  Started by 2 ex-Insightful colleagues

•  Sales Guy (BO, Cognos etc)

•  Techy Guy (S+, SAS, R etc)

•  Idea to deploy predictive analytics to business users

Why Mango?

•  Early awful ideas

•  DataStatz

•  Stats Entertainment

•  VizUStat

•  Stats2U

•  In the end, named after my colleagues cat

What we do?

R Training

Code Creation

Validation

Support

Consultants

What we do?

Consultants Developers

Analytic Application Development

Mango Key Industries

•  Mango work across sectors:

• Pharmaceuticals

• Mango Imaging

• Finance

• Energy

• Sensory

Why Build Analytic Applications on R?

Why Analytics?

•  Analytics can help people answer all sorts of questions •  I believe there is no company in the world today who

cannot benefit from analytics in some way •  More and more people are realising it

Who is a good driver? How do we win more games? What bonus should I pay?

Will someone like this? What are they likely to want? When might this break?

Why build Analytic Applications?

•  3 key reasons we see: •  To deploy analytical tools to decision makers •  To make an analysts life more efficient •  To add rigour to an analysts workflow

Deploying Analytics

•  Adding analytics into a business process can mean more informed decisions can be made

•  Complex analytics shouldn’t be attempted by non-analysts

•  Means there is a communication between the decision maker and the analyst

Deploying Analytics

•  If we build an application which … •  is easy for the decision maker to use •  contains the correct analysis to apply •  communicates analytical results in suitable manner

•  … this leads to some major benefits!

Benefits for the Analyst

Benefits for the Decision Maker

No need to wait for information

Can perform “what if” analysis

Decision not dependent on analyst availability

Less need to perform often-repetitive tasks

Comfortable that the “right” analysis is being run

Can get on with more strategic things?

Analytic Engine

User Interface

Data

Analytic Outputs

Data Storage

Analytic Code

Code Mgment

Analytic App Structure

Why build Analytic Applications on R?

Building applications requires installing analytic engine on desktops, servers, clusters, clouds R is license free

Building analytic applications involves integrating an analytic engine with other technologies (data sources, UI etc) R’s open nature means it can be readily integrated

Why build Analytic Applications on R?

We want a programmable engine so that it can be readily extended (i.e. no black boxes please) R can be extended by the developer as needed

We often want to be able to deploy new algorithms and techniques as they become available R is rapidly developed

Formal R Application Development

Formal R Development

•  Creating sophisticated analytic applications requires a formal development approach

•  This mostly means taking standard development practices and applying it to analytics

•  Mango’s formal R development procedures and structure has been evolving since its inception ~2004

Project Mgment Requirements

Behaviour Driven

Code Review

Review board

StatET

runit roxygen2

Continuous Integration

Issue Tracking Quality Manual

Dev Procedures R Coding Standards

mangoUtils

Knowledge Mgment

Project Mgment Requirements

Behaviour Driven

Code Review

Review board

StatET

runit roxygen2

Continuous Integration

Issue Tracking Quality Manual

Dev Procedures Coding Standards

mangoUtils

Knowledge Mgment

Project Mgment Requirements

Behaviour Driven

Code Review

Review board

StatET

testthat roxygen2

Continuous Integration

Issue Tracking Quality Manual

Dev Procedures R Coding Standards

mangoUtils

Knowledge Mgment

Case Studies

Case Studies

•  These are examples of applications we’ve built that use R in some way

•  We’re presented a range of information about each including: •  Business Reason for the application

•  Technical Approach

•  Some Technical Detail where applicable

•  Things that worked well / things that didn’t

Case Studies

•  Ranges from information we can fully disclose to only being able to say vague things about the customer

•  Only so much info we can give today – please see us after or contact us and we can step through things in more detail

Richard Pugh = rpugh@mango-solutions.com

Andy Nicholls = anicholls@mango-solutions.com

Chris Campbell = ccampbell@mango-solutions.com

Case Studies

•  PKPD Web Modelling Platform

•  M&S Workflow Platform

•  Non-Compartmental Analysis Application

•  Coffee Blend Optimisation Tool

•  Pipeline Corrosion Forecasting Application

•  Backtesting Application

CASE STUDY PKPD WEB PLATFORM

Case Study: PKPD Modelling Overview •  Pharmacokinetics-

pharmacodynamics (PKPD) is the study of the manner in which a drug transitions through the body and its impact on a target disease

•  PK is highly complex, involving sophisticated non-linear mixed effects modelling approaches

Case Study: PKPD Modelling Overview •  Modellers use “NONMEM” software in order to fit

these models

•  Inputs and outputs to NONMEM are a mixture of structured and unstructured textual files

•  R often used to analyse the outputs in order to assess model fit (see “xpose4” library)

Case Study: PKPD Modelling Overview •  PKPD is an evolving and exciting area, with

modellers needing flexibility and a variety of tools

•  However, being within life sciences, rigour around workflows is key in order to satisfy regulatory requirements

Case Study: PKPD Modelling The Challenge •  Build a modern modelling platform that provides rigour

whilst allowing the modellers the flexibility they need

•  Range of technical users from “everything is a shell script” to “which button do I click”

•  Execution of third party tools (NONMEM, R, SAS, PsN, …) in a controlled manner

•  Interface to generate reproducible graphics, tables and reports

Case Study: PKPD Modelling The “R” bit •  Where does R fit in?

• Many users use R and want to be able to develop scripts and execute them on an internal grid

• R used as the graphics engine to support the model evaluation and reporting processes

• Users want to be able to execute R interactively with objects in their project

The App

App Server

Execution Server(s)

MIF MIF Queue

Cloud

Grid

+ Others

+ Others

RPool Mgr

Case Study: PKPD Modelling What is a “Report Item Definition” •  Definition of a graph or table that can be executed

from Navigator

•  Consists of snippet of R code, options that may be presented to the user, required columns, and a few other bits

•  Can be used in a number of situations in the application

•  Originally XML then stored in Db (XML shown to give a feel for structure on next slide)

Report Options

Source Data

Command Definition

The App / RPool Manager

Data Graph Table Text

Data Item

Graph Item

Table Item

Text Item

Data Frame Graphics Table

Object Character

xml Method

xml Method

xml Method

xml Method

Execution Engine (Java)

Command Definitions

Command Results Ve

rsio

n Con

trol

Case Study: PKPD Modelling How are “RIDs” used? •  Created, managed by Super Users (under version control)

•  Called in a few places in the application:

•  Directly (create this graph with this data)

•  In “Run Views” (reports)

•  In “Comparison Views” (reports that compare models)

•  In “Template Reports” (tagged docx files)

Case Study: PKPD Modelling Outcome •  The app in general was a big success

•  The “R” part was created as a separate service that we have since reused in a number of other applications (e.g. Lloyds Risk Platform!)

•  Shame that regulatory rules forced some design which we’re now building alternatives too

•  Next: interactive graphical presentation

CASE STUDY M&S WORKFLOW PLATFORM

Case Study: M&S Workflow Platform Overview •  Exciting project for major pharmaceutical company

•  Possibly the closest we’ve come to deploying an analysts workflow in a scalable platform

•  Hundreds of pre-clinical (animal) studies are run by a team of ~400 scientists

•  Analysis performed by roughly 15 advanced modellers

•  Outcome: most studies not analysed!

Case Study: M&S Workflow Platform The Challenge •  Idea to create a truly scalable platform to allow bench

scientists to run their own analysis

•  Modeller publishes an analysis “protocol” containing analysis paths, code, and support documentation

•  Desktop application pulls from central set of protocols and “derives” the interface which is presented to the user

•  Modelling can put in checks to ensure things look right (e.g. data is of right format, model fit is particularly poor but user seems keep to create predictions from it)

Case Study: M&S Workflow Platform The Solution •  Eclipse RCP application executing R and NONMEM scripts on

an internal LSF grid, with protocols and code held in SVN

•  Generated workflow “protocol” definition (XML) detailing possible paths in a step, linked to R scripts and NONMEM model code with corresponding dialog

•  Built “Protocol Developer” Eclipse interface onto repository

•  RCP application derives analysis paths, UI, options and commentary to guide the end user

Protocol Metadata

Workflow

Analysis Step Data Check Step Commentary

R Script

R Script

NM Model File

Options Options

Modeller Scientist

NONMEM

NONMEM

LSF Grid Protocol Server

File System

Possible Models Derived Options Commentary

Case Study: M&S Workflow Platform How did it go •  Technical solution was very strong and applicable to

other areas

•  RCP good technology, but steep learning curve

•  Testing was complex

•  Agile project – pros and cons

•  Ultimately, not deployed (site closure)

CASE STUDY NON-COMPARTMENTAL ANALYSIS

RapidNCA, the non-compartmental analysis workflow tool •  Need for RapidNCA

•  Using .NET

•  RapidNCA Structure

•  Code Quality

•  Connections with R.NET

•  Complete & Deploy RapidNCA

Need for RapidNCA

•  Customer needed to send monthly reports to dozens of trial centres

•  Small team, so time limited

•  Predefined non-compartmental analysis

•  Standardized report

Using .NET What is .NET?

•  Object-oriented environment to develop applications

•  Safe execution environment

•  Choice of programming languages

•  Framework consisting of:

•  runtime

•  class library

•  Developed with Visual Studio

Using .NET Visual Studio

•  A graphical programming tool (IDE)

•  Visual Studio Express - free version

Using .NET Choice of languages

•  C# is the main one

•  F# is a functional language (similar concepts to OCaml)

•  XAML (a Microsoft declarative XML language) for interactive graphics

•  C++/CLI useful for legacy and bespoke parallel processing (including GPGPU)

•  Other possibilities...

•  Vb.Net is very like C# (no advantage over it)

•  Third parties have added languages to the CLI platform

Using .NET “Ajar Source” Platform

•  Not exactly open source, but…

•  Most CLI third party languages are open

•  C# and VB.Net are not, but many open source projects based on them

•  Microsoft have made F# open source

•  Compiler is free

•  Other editors / IDEs are available

Using .NET Performance

•  Performance is very good

•  On graphics (millions of data points will plot with ease and zoom smoothly)

•  Computation is fast enough in C#, calling R adds little overhead

•  Standard Maths library is limited; third parties and MS maths for “drawing” are better

•  Data parallel computation is possible on the desktop (GPGPU)

•  F# provides further “big data” capabilities

User Interface

Data

Analytic Outputs

Data Storage

Analytic Code

Code Mgment

Analytic Engine

Data Service

RapidNCA Structure

RapidNCA Structure MangoNca Analytic Code

Analyse Element

Do Analysis

Get Analysis

Unit Tests

Data Checks

RapidNCA Structure MangoNca Analytic Code

Code Quality Unit Tests

•  Ensure product works!

•  User/Customer/Payer trust

•  Ease of maintenance/extension

Code Quality Run Code, Check Output

•  Working Cases

• 

> test1 <- ncaAnalysis(Conc = c(4, 9, 8, 6, 4:1, 1), + Time = 0:8, Dose = 100, Dof = 2) > checkEquals(test1[1, "ROutput_adjr2"], 0.9714937901, + tol = 1e-8) [1] TRUE

> require(RUnit) > # there are other automated test packages!

Code Quality Error Case Unit Tests

•  Use try

• 

•  Handled Error Cases

> test7 <- try(AUCLast(Conc = 1:10, Time = 9:0), + silent = TRUE) > checkEquals(test7, + "Error in checkOrderedVector(Time, ... ") [1] TRUE

> test26 <- ncaAnalysis(Conc = c(4, 9, 8, 6, 4:1, 1), + Time = 0:8, Dof = 1) > checkEquals(test26[, "ROutput_Error"], + "Error in checkSingleNumeric(Dose, ... ") [1] TRUE

Connections with R.NET

•  What will be provided to R?

•  What will be returned from R?

•  What happens if something goes wrong?

Connections with R.NET Using the R Service

•  R.NET allows R calls to be submitted to an R service

•  R.NET connects to R down to Expression level

•  Data from return objects passed back into .NET

Connections with R.NET Data Checks

•  Function may be passed data outside its anticipated structure

> checkOrderedVector(c(0, 1, 3, 2, 4), + description = "Time") Error in checkOrderedVector(c(0, 1, 3, 2, 4), description = "Time") : Error: Time is not ordered. Actual value is 0 1 3 2 4 >

Connections with R.NET Data Checks

•  The tool expects a certain return object

•  An error in an R call should be trapped by the communicating function

•  Return object passed as normal

•  An error checking element of the return object can report information about the error

> check01 <- try(checkOrderedVector(Time, + description = "Time"), silent = TRUE) > if (is(check01, "try-error")) { return(object) }

Connections with R.NET

_pluginsManager = new RPluginManager(PluginLocation, RLocation); _pluginsManager.SetActivePlugin(); _session = _pluginsManager.GetSession(); bool sessionOk = _pluginsManager.TryMakeSession(out _session);

•  R is efficiently accessed, via R.Net (as pictured in Visual Studio) via a Plugin (as above)

Connections with R.NET

User Interface

Data

Analytic Outputs

Data Storage

Analytic Code

Code Mgment

Analytic Engine

Data Service

R.NET

Analysis Display

Get PK Params

Data Service

Dialog Service

App Logger

Status Bar Service

App Config

Mgment

Data Importers

Project Wizard

Validators

Receive R Output

Create R Expressns

Connections with R.NET .NET Data Service

R.NET

Connections with R.NET Using the framework

_pluginsManager = new RPluginManager(PluginLocation, RLocation); _pluginsManager.SetActivePlugin(); _session = _pluginsManager.GetSession(); bool sessionOk = _pluginsManager.TryMakeSession(out _session);

_session.SetNumericSymbol("TimePtVector", CheckTimePointData(toAnalyse)); _session.SetNumericSymbol("ConcVector", CheckConcentrationPointData(toAnalyse));

var evalString = string.Format("ncaAnalysis(TimePtVector, ConcVector, …

MathEngineDataRowDto<double> ncaGetBack = _session.PerformNumericEvaluation(evalString, "ROutput_Error"); _lastErrors = ncaGetBack.ErrorStrings;

_session.FlushConsole(); _pluginsManager.RelinquishSession();

Complete & Deploy RapidNCA

•  Can users understand how to use tool?

•  How confident are we in tool output?

•  On-going code review

•  Independent test team

•  Installation Qualification

•  Operational Qualification

•  Performance Qualification

Deploy Tool

Data Import

Map Variables

Review Analysis

Review Grouping

Generate Report

Select Report Type

Add Group Comments

View Report

Conclusions

•  Great graphical interfaces can be built using .NET

•  Intuitive interactive features are available

•  R.NET allows R analysis to be accessed as a service

•  Good coding practice will ensure application is robust

•  Work on a well engineered framework will be rewarded with desktop solutions created at high speed

CASE STUDY COFFEE BLEND OPTIMIZATION

Company Background

•  A global chocolatier, biscuit baker, candy maker and maker of gum.

Business /Technical Situation

•  The client was using a desktop SPLUS application to simulate and optimise coffee blends for their manufacturing teams

•  Hugely successful application saving the company $millions

•  They wanted to make improvements and expand the usage beyond Global Statistics and beyond coffee

•  Also keen to remove the license fee

Application Workflow

Import Data from Excel

Graphical Visualisations

Export Data

Run Blend Optimiser

Simulate Blends

Audit Log

System Architecture

Functions for GUI

Functions for Analysis

R Package

Optimizer

Data Import

Data Export

Approach •  Development phase split into three separate pieces:

•  Code conversion

•  GUI creation

•  Development and integration of a new optimiser

•  Each required the generation of unit and system tests and appropriate documentation, including help files

•  Design specifications captured prior to development

•  Project estimated at c90 man days over 3 months

Creation of new GUI

GUI Choices Some R/R-based technologies we could have used...

•  tcltk is R’s ‘recommended’ menu builder

•  Glade, RGtk2

•  gWidgets

•  rpanel •  Deducer •  manipulate (Rstudio)

•  ...

GUI Choices

Other options:

•  Choice is almost limitless

•  Often they require a knowledge of other languages such as Java or C

•  Possibly warrants a standalone talk...

Creation of a New GUI using RGtk2

•  RGtk2 adapter for R of the GTK+ engine

• Gimp Toolkit

•  Glade can be used to trial new features

•  GTK allows for automated testing of the GUI

• Huge time saving

Code Conversion

Mango took a test-based approach for the code conversion (RUnit)

•  Allows for automated testing in future revisions

•  Simple PASS/FAIL reporting

•  SPLUS knowledge not required for R code development

Optimization

•  The original SPLUS application used the SPLUS NuOpt optimizer

•  R NuOpt exists but only on license

•  Mango used an open source optimiser that we integrated into the R GUI

•  Mango implemented a ‘quick run’ option to allow quick comparisons with the simulation piece

Primary Benefits

•  New departments are now benefitting from the application

•  The application is now in the hands of the manufacturing teams, reducing the burden on Global Statistics

•  Test-based approach facilitates future development of the application

CASE STUDY PIPELINE CORROSION APP

Background

•  One of the biggest companies in the world with thousands of staff

•  Oilfield Exploration Team based in the UK but with responsibility for complex exploration areas

• Alaska, shale fields etc

Business Situation

•  Thousands of miles of pipeline corroding in freezing, isolated areas

•  How do you choose how often to inspect them?

•  The cost of a leak can run into many billions of £s

Technical Situation

•  Customer Team were analysing data using S-PLUS Insightful Miner with many non-analytical workarounds

•  Process was messy and took a long time to run

System Architecture •  This piece is one of several in a continuous workflow

•  All information is fed back into the database

Functions for GUI

Functions for Analysis

R Package

Access Database

General Workflow

Read

Write

Approach •  Consulting engagement to improve programming

techniques and statistical methodology

•  Create an R package for the code

•  Construct a GUI in order to deploy to non-technical users on the frontline

An Interesting Challenge: Converting S-Plus Code to R

This is Easy, Right?

Some (true?) statements:

•  R can be considered as a different implementation of S

•  There are some important differences, but much code written for S runs unaltered under R

Discuss...

Source: www.r-project.org

Considerations

S+ applications can generally be split into two pieces:

•  An underlying library of code

•  A set of functions defining the menu system and help pages

Approach

There are essentially two approaches to code conversion:

•  Direct Conversion

•  Test-based Conversion

Direct Conversion

•  Requires knowledge of both languages (stdev vs sd)

•  Relatively quick to achieve

•  Difficult to prove the new code does what the old code did

Test-based Conversion

•  Generating unit tests in S+ requires some S+ knowledge

•  Takes some time to generate and document tests but better in the long-run

•  Unit tests give a definitive PASS/FAIL result

•  Can often be automated

Code Conversion Challenges

•  The application upgrade usually coincides with an operating system upgrade

•  Windows (or other) version and R version need to be determined in advance

•  It is almost guaranteed that the new system will produce different results for the same test data!

What is “different”?

•  Often this is simply rounding

• Still require agreement on precision: 0.049782 vs 0.050436

•  If simulation is involved this can be VERY difficult to define!!!

•  Appearance of graphics may also differ

Other Challenges

As the business owner I want to use the opportunity to improve the application:

•  New menu items

•  New functionality

•  Modifications to existing functionality

All of these require careful planning

Primary Benefits for Customer

•  Rationalised code base means the analysis is quicker and extensible by end-users

•  Construction of a front-end has enable rollout to users on the font-line in Alaska

•  Conversion to R has removed license cost

CASE STUDY BACKTESTING APP FOR HEDGE FUND

Case Study: Backtesting App Overview •  Backtesting has a key role to play in the testing of

automated trading strategies

•  Asked by a Hedge Fund Manager to build for his team of users (who love Excel)

•  Mango were asked to build a backtesting platform that was more sophisticated that what was on offer from other vendors

•  Sorry that the details may be occasionally sketchy in this section

Case Study: Backtesting App The Challenge •  Key parts of the challenge included:

•  Integration with standard finance data streams

•  Advanced portfolio optimisation

•  Flexibility to define automated strategy

•  Transaction-cost based benefit analysis

•  Leverage of financial hurdle

•  ARCH-style error incorporation

•  Advanced reporting

Alpha Storage

Data Storage

Data Flow

.NET Interface

RdotNet

C Interface!

.Rda Files

How I learnt apply functions!! Some hacky code here …

Case Study: Backtesting App The Outcome •  Very successful hedge fund

•  Convinced the users to use R – UI dropped!

The R Community

IP Considerations

•  IP based on R includes: •  New libraries & code

•  New scripts

•  Mango attempt to open source (with client permission) any “R-side” generic functionality

•  Also feedback and assist library authors

User Interface

Analytic Code

New R Libraries

Great Example

•  MSToolkit library built for Pfizer

•  Funded by Pfizer, built by Mango

•  Released as open source library

•  Since extended by other companies

R Community

•  Contribute code where allowed/useful

•  Sponsor R conferences and events

•  Provide free training courses / webinars

•  Organise and fund many R user groups (LondonR, BaselR, ZurichR, ShanghaiR, NewJerseyR, …)

The End!

Summary

•  Thank you for the invitation •  Hope the discussion was useful

•  We could only cover certain amount of detail in time, so ask us for more if interested!

Andy Nicholls anicholls@mango-solutions.com

Chris Campbell ccampbell@mango-solutions.com

Richard Pugh rpugh@mango-solutions.com