A select few posts from 2013

Rstudio's Shiny enabled me to try out ideas on the Web.

• WeatherCompare
• shinySketch
• EscherPlot
• MovieRatings

** Creating Mazes


The most visitors came from these two posts:

• The hottest areas in CS for 2013 - per Google Research grants
• Anand-vs-Carlsen: What the numbers tell?

Happy 2014, everyone.

Ram

Anand versus Carlsen - Chennai 2013 - What can we expect in November?

This year, the world chess championship will be played between Vishwanathan Anand and 22-yo Magnus Carlsen, in Chennai, India from the 9th to the 28th of November. The passions are sure to run strong. Both GMs have ardent supporters. Carlsen is in a dreamlike form, and Anand has the experience and the home field advantage. But what do the numbers say?

Let's use R to look at some data and see what we can infer.

The data comes from chessgames.com.  I took the raw data and created two new columns to facilitate my analysis. 1. A column called Anand.White (1 or 0) and 2. a column called Anand.won (which is a factor with 3 values: 0, Draw, 1). The cleaned csv files can be found here.

1. Lifetime tally

This is always a good place to start. We have data for a total of 62 games. Anand has a slight lead on this count, with 3 more wins than Carlsen. (Not distinguishing between rapid and standard games here.)

In R, we can simply run the table() command on the Anand.won column.

 Loss  Win   Draw
  11   14     37 

2. How has each GM grown in strength

We can use ELO ratings since 2000 to see how both GMs have performed over time. Anand, of course, has been in the top 5 in the world for the past two decades, pretty much since Carlsen was born! But the visual showing Magnus' meteoric rise is quite striking. (The data comes from FIDE.com and can be found here.)

3. Win-Loss-Draw record by Year

Let's say we want to look, year by year, how the two GMs have fared against each other. R has this great package called "plyr" which is tailor made for these kinds of "Split-Apply-Combine" type analyses. We are splitting the data by year, and combining based on win-loss, and plotting the tallies. The plotting package ggplot plays well with the output of plyr.

Once we do the plotting, we get a sense of what has been happening. In the early 2000s, Anand had a much higher share of wins. Overall the number of draws has gone up over the years. But Carlsen has had the upper hand the last year or two. (Of course, the number of games is too small, and we should be careful about "inferring" when the data is this tiny.) That said, we could make a strong the case that Carlsen has the momentum going for him.

4. Choice of Openings

Finally, we know that both GMs are holed up somewhere with their team of seconds and coaches, preparing. What do these experts prepare? A good majority of the times they are preparing opening surprises to spring on their opponents. They are studying each others' games looking for weaknesses. By looking at how their choice of openings helped them in the past, we can make a broad guess about what they might go.

R allows us to slice the data by their choice of openings, and we can see how they fared.

So we can expect that Anand will favor the openings that have more "green" (wins) for him, while Carlsen will try to play the openings that have been "red" (losses) for Anand.

By this logic, we can expect Anand to opt for the Queen's Gambit Declined-semi-slav(D47), the Ruy Lopez, closed (C96)  or the Sicilian closed (B23). Magnus will be trying to steer the game towards the English (A20) or the Benko (A58) which are slightly more unorthodox, but have served him well against Anand. The expansions behind the ECO list can be found here.

Of course, there will always be surprises. (This is where it all gets game-theoretical. If only it were this easy to predict...) And that's why we should watch what unfolds this November.

The R code used can be found here.

Ram

Using R to visualize Karpov-vs-Kasparov Lifetime winner-take-all tally

The Karpov vs. Kasparov rivalry holds a special place in the chess world.

The idea behind this analysis is simple. If we take their lifetime games, plot the wins, what would it look like? We introduce one twist -- we'll be plotting the "winner-take-all" tallies, meaning that for every year, every five years, and every decade, we declare one person to be the 'winner'.

First, a note of caution: "Winner-take-all" type analyses lose a lot of information due to the roll-up. Whether a GM wins by 1 game, or a dozen games in a given year, he still gets only one "win".

At the outset, I must mention that this is NOT a chess exercise. I am ignoring the colors (whether each player had White or Black pieces) and even more egregious, I don't differentiate between standard and rapid games, or exhibition games. Time controls are ignored, as are openings.

This is a visualization exercise, and the idea is to see how it all looks when plotted.

I scraped the data from chessgames.com - where they have 201 games that the two have played. (I cleaned up the data and the csv file is available in github for anyone who wants to do their own analysis.) I use plyr to aggregate the data, and ggplot for the visualization. I wanted to try out this "pianogram" type visualization, where each plot looks like piano-keys.

Let's get the basics out of the way:

201 games - 138 draws,  37 wins for Kasparov, 26 for Karpov

Overall, Kasparov pretty much dominated Karpov. But how are these wins and losses spread across time? The two played for a little over 30 years.

The Winner-Take-All Method
In any given time period, say 1990, there can be 4 possible outcomes:
No games played, Equal number of wins, Karpov won, or Kasparov won. (If both players had the exact same number of wins in a given time period, we label that a Draw.)

Thanks to the 'plyr' package and ggplot, we can calculate the by-year, "5-year Winner" for each half-decade, and also the decade-wise winners by writing one function, and calling it with ddply.


So here's what the Yearly-Winner-Take-All looks like:

Let's plot the half-decade and decade plots. Again, note that only one GM is declared the winner for the entire decade, no matter what the difference in the scores are.

Now, we can put it all together, in one graph.

As a very quick summary, we can see that Karpov started out strong, the entire 80's was a draw, and then Kasparov took over.

The complete R code to reproduce this analysis is available in this gist, along with the data-file in CSV file.

What are the hottest areas for CS Research? (Based on Google Research 2013)

What are some of the hottest areas of research in Computer Science at the moment (August 2013)? And at which universities is this research being carried out?
The answers are subjective by definition, but looking at the numbers behind the Google Research awards announced yesterday can provide some quick insights. Going by the grants as a proxy for what are where are the current hottest areas of research, here's what we get:

A total of 105 grants were awarded. 81 universities in all got the awards, in 19 different research areas.

Here are the top research areas:

As far as institutions go, MIT, Georgia Tech and CMU got 5 grants each, with Cornell and Standford getting 3 each.

The R code I used to generate this can be found here. In case anyone is interested in performing their own analysis, I've also included the CSV data file.

Using R and Shiny to create “Art”

One big strength of packages like shiny is the ability to easily vary parameters and view the results, especially in plots.

So here’s a small shiny app that I created to learn about reactivity, while also having fun.



The idea is simple. Vary many aspects of geom_segments in ggplot, and see what emerges. The things that I played with are canvas size, line origin and destination, line length, the angle and the colors.

Because it is art, I made the background black.

There were many experiments that didn’t appeal to me aesthetically. Others seemed very repetitive. The ones that seemed okay, I kept.

Take a look at the shiny app ShinySketch.

here:

The code can be found on github.

Using R and Integer Programming to find solutions to FlowFree game boards

Using R and Integer Programming to find solutions to FlowFree game boards

 

What is FlowFree?
A popular game (iOS/Android) on a square board with simple rules. As the website states: Connect matching colors with pipes to create a flow. Pair all colors, and cover the entire board to solve each puzzle in Flow Free.

If you play a few games, you will be able to follow this post better. The games get progressively harder as you increase the board size.

 The solutions to the various levels are available on the web, but it is fun to use R (and the lpSolveAPI package) to come up with solutions ourselves.

So let's say that the board is made up on n "cells" (small squares) on each side. A pipe can enter a cell from the center of any side. So each cell has 4 edges.

 Figure: A cell with 4 edges meeting at the center

Terminal cells - Those cells where a colored-strand begins or ends. (Only one edge can be on in that cell.)

The 'rules' that we need to incorporate into our IP model
* The lines cannot cross.
* All the cells (squares) must be covered
* The colored lines must begin and end where they are shown

Decision variables

The problem comes down to deciding which edges to turn on, and which ones to leave off. An edge has a color and a cell. X_cke is 1 if edge e of color k in cell c is valid. 0 otherwise.

The constraints in English

1. Cell Cover Constraints: Every cell has to have 2 edges that turned on. (One to come in, one to exit) 
1. Special case: Terminal cells have only one edge that can be 1. All other edges are 0. 
2. An Edge can have only one color (LimitEdge constraints) 
3. Horizontal Connectivity - If two cells are horizontal neighbors (side by side), then the value of the east edge of the left cell must be the same as the west edge of the cell on the right. 
4. Vertical Connectivity - If two cells are vertical neighbors (one on top of the other), then the value of the downward edge of the top cell must be the same as the upward edge of the bottom cell 
5. Boundary Edges - Certain edges for the cells in the boundary of the board are made zero. (For example, in the 4 corner cells, at most 2 edges can be 1.) 
6. Single color in Terminal Cells - Each terminal cell has a pre-determined color given to us as input, so we can set of the edge varibles of every other color to be zero.
7.  Same color per cell & Pick 1 constraints. We set 0/1 variables and make sure that if a color is one inside a cell, ONLY that color edges are allowed in that cell.  (These are dense constraints and can make solution times explode for larger grids.)

Load one specific Puzzle

All of the work of  creating the matrix, the right hand side are done by 3 functions - init(), populate.Amatrix() and defineIP(). [See github for the code] Now, let's initialize an empty data frame, based on number of constraints and variables that the IP will have.

Getting the problem ready for LPSolve

After using lpSolveAPI's solve() function, we plot the solution using ggplot2 to recreate the board and show the pipes in them.

Let's try one problem:

>plotProb(terminal.cells, colorpalette)

 

 

Here are the 8 terminal cells for this problem:

> terminal.cells
  X Y color palette tcell
1 1 2     1   green     6
2 5 4     1   green    20
3 1 1     2    blue     1
4 2 4     2    blue    17
5 3 4     3  yellow    18
6 4 2     3  yellow     9
7 3 1     4     red     3
8 4 4     4     red    19

We create the linear program, and solve it using lpSolve

Let's see how many edges have non-zero values:

> lpff <- make.lp(nrow=n.row, ncol=n.col)
> defineIP()
> lpff

Model name:
  a linear program with 500 decision variables and 458 constraints
> solve(lpff)
> sol <- get.variables(lpff)

> sum(unlist(sol[1:num.edges]))
[1] 42

> plotSol(terminal.cells, colorpalette)
 

 produces the solution:

 

And here's the solution when I tried an 8x8 (Level 17) grid:

Note: Because of same-color and pick-1 constraints, this problem explodes in size as n (the size of the grid) increases. There are however, ways to address it, using relaxation.

Future Improvements: The addition of the pick-1 and the same-color constraints spoil the structure of the A-matrix and increase the solution time. These can be made optional, in which case the solution will require manual intervention.



You can download the full R code from github.

Useful Links:
1. Linear programming in R: an lpSolveAPI example (fishyoperations Blog)
2. Using lpSolve from R (I started with their working example)

R Beginners - Plotting Locations on to a World Map

This post is targeted at those who are just getting started plotting on maps using R.

The relevant libraries are: maps, ggplot2, ggmap, and maptools. Make sure you install them.

The Problem

Let's take a fairly simple use case: We have a few points on the globe (say Cities) that we want to mark on the map.

The ideal and natural choice for this would be David Kahle's ggmap package. Except that there is a catch. ggmap doesn't handle extreme latitudes very well. If you are really keen on using ggmap, you can do it by following the technique outlined in this StackOverflow response.
 If ggmap is not mandatory, there are simpler ways to do the same.

First, let's set up our problem. We'll take 5 cities and plot them on a world map.

Method 1: Using the maps Package

This results in:

Which might be enough. However, if you take the few extra steps to plot using ggplot, you will have much greater control for what you want to do subsequently.

Method 2: Plotting on a World Map using ggplot 

 

This results in:

R - Defining Your Own Color schemes for HeatMaps

This post is intended at those who are beginners at R, and is inspired by a small post in Martin's bioblog.

First, we plot a "correlation heatmap" using the same logic that Martin uses. In our example, let's use the Movies dataset that comes with ggplot2.

We take the 6 genre columns, and we can compute the correlation matrix for those 6 columns.
Here's what the matrix looks like:

> cor(movieGenres) # 6x6 cor matrix
                  Action   Animation      Comedy        Drama
Action       1.000000000 -0.05443315 -0.08288728  0.007760094
Animation   -0.054433153  1.00000000  0.17967294 -0.179155441
Comedy      -0.082887284  0.17967294  1.00000000 -0.255784957
Drama        0.007760094 -0.17915544 -0.25578496  1.000000000
Documentary -0.069487718 -0.05204238 -0.14083580 -0.173443622
Romance     -0.023355368 -0.06637362  0.10986485  0.103545195
            Documentary     Romance
Action      -0.06948772 -0.02335537
Animation   -0.05204238 -0.06637362
Comedy      -0.14083580  0.10986485
Drama       -0.17344362  0.10354520
Documentary  1.00000000 -0.07157792
Romance     -0.07157792  1.00000000

When we plot with the default colors we get:

It is difficult to see the details in the tiles. Now, if you want to better control the colors, you can use the handy colorRampPalette() function and combine that with scale_fill_gradient2.
Let's say that we want "red" colors for negative correlations and "green" for positives.
(We can gray out the 1 along the diagonal.)

Doing this produces:

If there are values close to 1 or to -1, those will pop out visually. Values close to 0 are a lot more muted.

Hope that helps someone.

References: Using R: Correlation Heatmap with ggplot2

R - Simple Recursive XML Parsing

This is intended for those who are starting out in R and interested in parsing an XML document recursively. It uses DT Lang's XML package.

If you want to just read certain types of nodes, then XPATH is great. This document by DT Lang is perfect for that.

However, if you want to read the whole document, then you have to recursively visit every node. Here's the way I ended up doing it. The generic function visitNode could be useful if you are just starting out reading XML in R.

The full code, along with a sample XML file to test it is here.

Simulating Allele Counts in a population using R

This post is inspired by the Week 7 lectures of the Coursera course "Introduction to Genetics and Evolution" (I highly recommend this course for anyone interested in genetics, BTW.) Professor Noor uses a Univ Washington software called AlleleA1 for trying out scenarios.

We can just as well use R to get an intuitive feel for how Alleles and Genotypes propagate or die out in populations.

Basic Scenario

There are N individuals in an isolated island. Say, we are interested in two specific Alleles (Big "A", or small "a"). This in turn means that they can have 3 types of genotypes: AA, Aa or aa. The individuals mate in pairs, and produce two offspring and die out. (Thus the total population remains the same generation after generation.)
The genotype of the offspring depends on those of the parents. A 'gamete' has only one parental allele, depending on what the parent's genotype was. AA type parent can only product gamete type A, aa parent can only produce gamete type a, but Aa can produce either type of gamete.

A Punnett square of parents gametes to offspring's genotypes. 

  | A  | a

 ---------------- 

A | AA | Aa 

a | Aa | aa 

With these simple rules, we can use R Simulation scripts to observe what happens to the Allele Frequencies over generations. (The goal here is to learn to use R for Monte Carlo simulations.)

Writing the R Script from scratch

 I toyed around with the idea of using character strings for the genotypes and the alleles. But then I realized that are only three types and I could just as easily represent them with the numbers 1, 2, 3 as a simple R vector.


With that done, we can write very simple functions for the procreation process.
With these useful functions, we can take one generation and produce another, 2 offspring for each set of 2 parents.
Putting it all together to generate multiple trials:

We also need to compute the Allele counts for each generation, and for plotting I use ggplot.



Using this simple Monte Carlo "toy" we can develop quite a bit of intuition.



For small starting populations, either the big A or the small a allele takes over the entire population fairly quickly. Given large enough number of generations, invariably one of the alleles gets wiped out.














As one example, we can see that even a small increase in the probability of Allele A to be 0.53 (up from 0.5) makes it take over quite dramatically.

Conversely, setting it to any value under 0.5 means that the Big A allele gets wiped out of the entire population.

The entire R script can be found here. You can download the code and try playing with various starting scenarios, changing the starting population counts, generations and probabilities.


References:

1. Coursera.org (Introduction to Genetics and Evolution by Md. Noor, Week 7 lectures) 
2. AlleleA1 software at Univ Washington

Simulating Population Growth in Cities Using R

R is great for anyone who wants to get started on learning Simulation. (Both Discrete Event or Agent-based, with stochastic elements in the process.)

This post is inspired by Matt Asher's "quick-and-dirty" R simulation work on Population Growth. Matt uses it to create aRt. I felt that his core idea provided a very good framework to introduce various concepts of simulation.

The Basic Idea behind this Simulation

We take an area (it could be a country, state, city) that is made up of units (cells) that can be settled (occupied).
We allow the initial 'pioneers' to settle first. (Seeding the cells)
We then establish some rules for future residents to settle. (Settling Rules)
Plot the area, and collect some stats to study the growth of the population under different starting conditions and different settling rules.

Pseudo-Code for the Simulation

Define parameters: Area Width, Height

Number of Pioneers P and future settlers S

for each pioneer p{ 

  settle p according to seeding rules in a valid cell in the area

}

for each settler S{

  Find an unoccupied Cell.

  Settle S in that cell, if the settling criteria permits it

}

Print Iterations Stats
Print out the Plot of the Area

A Few Changes from Matt's R code

1. I switched from his matrix to a dataframe. I also switched from his image to ggplot.
2. Delineated the initial parameters, seeding and settling clearly so that anyone who downloads the code can experiment
3. Created a few different Seeding Functions for the pioneers (beyond random seeding)
4. Matt allowed settling if the new cell was adjacent to an occupied cell. I played around with a few different population settling functions.  (rectilinear, diagonal etc).
5. Added a function to collect a few statistics regarding population growth
6. Ran the simulation a few 100 times (without plotting) to compare different growth and settling schemes

Color Coding the Population

• Black cells : Unoccupied
• Blue cells: Pioneers
• Orange cells: Settlers who found a cell in 1-2 attempts
• Yellow cells: Settlers who found a cell after more than 2 scouting attempts

The full R code I used for these R can be found here.

Here are some runs with a 30x30 area.

Here's what one run looked like. (Blue cells are the initial pioneers. The other colors represent the number of steps taken for them to find a home cell.)

Trying Out Different Seeding Schemes (for the pioneers)

As opposed to starting out with randomly strewn cells, what if we started out with a city downtown (a densely populated central rectangle) and let the settlers follow them?



I tried a rectangluar ring (annular cells for pioneers) starting point to see how the population grew inside and outside. So I used this seeding function:

One final seeding scheme I tried was to have two thick bands (columns of pioneer cells) and to let the population grow from there. Seeding function:

 

 

 

 

 

 

Different Population Growth Schemes

The default settling rule that Matt Asher uses is a very good one, one that I like. If a given cell is adjacent to any occupied cell (one of its 8 neighbors) then it becomes a valid cell to settle down in.

Just for experimentation, I tried a few other schemes.

1. Rectilinear Growth

Rule: You can only settle down in the 4 cells that are due N, S, East or West of an occupied cell.

2. Diagonal Growth

Rule: A settler can occupy any cell that is diagonally adjacent to another occupied cell.

The code can be found here.

 

 

 

 

 

 

Collecting Stats

Now that we have all the functionality that we wanted, it is time to start collecting some statistics for each run ('replication'). There are many possibilities for what we could collect in a population simulation.
Here are a couple:

• How many settlers found a 'home cell?'
• How many scouting attempts did each settler make before they found a home?

Making Multiple Replications

 

>st
  Iter FoundHome NumSettlers  Percent
1    1       243         350 69.42857
2    2       214         350 61.14286
3    3       213         350 60.85714
4    4       258         350 73.71429
5    5       235         350 67.14286

The full R code I used for these runs can be found here. Feel free to give it a spin and try out your own settling/growth schemes.
Reference: Matt Asher's post on his StatisticsBlog