ReferenceClasses

Rob Bennetto and Eckhard Briedenhann
2019-04-06

Distributor of Icepack services
Consulting and Ops Research in Africa
Insurance and Logstics

$plot of chunk unnamed-chunk-1$

SatRday

What do we wear?
- Normal clothes?
- Formal clothes?
- Zebra suit?
How do we fit it in with all of our other responsibilities?
- Will there be enough time?
- How long will it take?
- What do we do first?
- Create presentation?
- Code up something interesting and hope the presentation self-materialises?
- Go find a corner to cry in?

plot of chunk unnamed-chunk-2

How do we over engineer this?

What do we need?

A way to represent a state of being
A way to represent the transitions between states as well as their attributes (time, awesomeness-factor, etc.)
Algorithms to calculate the the optimal set of state to traverse

plot of chunk unnamed-chunk-3

Graph?

plot of chunk unnamed-chunk-5

Process:

Create states
Assign weights to the transitions
And calculate metrics related to graph

Rob's face

plot of chunk unnamed-chunk-6

DAGs

Directed Acyclic Graphs (aka DAG)

Nodes (Vertices)
Directed Edges ( pairs of nodes)
No Cycles

plot of chunk unnamed-chunk-7

Representation problem

Nodes:

nodes <- list(n1 = list(id = 1),
            n2 = list(id = 2),
            n3 = list(id = 3), 
            n4 = list(id = 4))

Helper functions:

getNodeID <- function(node){
  paste0("n",node$id)
}

Vectorised Graph:

graph <- data.frame(From = c(1,1,2,3),
                    To = c(2,3,4,4))

List Graph:

graph <- list()
graph[[getNodeID(n1)]] <- list(n2,n3)
graph[[getNodeID(n2)]] <- list(n4)
graph[[getNodeID(n3)]] <- list(n4)

$n1
$n1[[1]]
$n1[[1]]$id
[1] 2


$n1[[2]]
$n1[[2]]$id
[1] 4



$n2
$n2[[1]]
$n2[[1]]$id
[1] 4



$n4
$n4[[1]]
$n4[[1]]$id
[1] 4

plot of chunk unnamed-chunk-15

Why does this smell funny?

No typing - live in the wild west
No standard - reinventing the wheel one line at a time
No abstraction - very verbose with a lot of index hunting

Why does this smell funny?

No typing - live in the wild west
No standard - reinventing the wheel one line at a time
No abstraction - very verbose with a lot of index hunting

plot of chunk unnamed-chunk-16

C++ World

Object Orientated Programming:

Node Class:

// Vector object
class Node {
  public:
    Node(int id):id(id){}
    int id;
};

Graph Class:


class DAG {
  public:

  DAG(int size) : adjList(size){}

  void addEdge(Node* from, Node* to){
    adjList[from->id].push_back(to);
  }
  std::vector<std::vector<*Node>> adjList; 
};

In practice:


// Init nodes
Node n1(1); 
Node n2(2); 
Node n3(3); 

// 1 --> 2 --> 3
graph.AddEdge(&n1, &n2);
graph.AddEdge(&n2, &n3);

Object Orientated Programming in R?

plot of chunk unnamed-chunk-17

Object Orientated Programming in R

S3 Classes:

n1 <- list(id = 1)

class(n1) <- "node"

print.node <- function(obj){

  cat("Class:",class(obj),
      "\nID:",obj$id)
}
print(n1)

Class: node 
ID: 1

Benefits:

At least there is some kind of structure
Class specific functions might be useful

Overall: 5/10

Object Orientated Programming in R

S4 Classes:

node <- setClass("node", slots=list(id="numeric"))

n1 <- node(id = 1)

print.node <- function(obj){

  cat("Class:",class(obj),
      "\nID:",obj@id)
}
print(n1)

Class: node 
ID: 1

Benefits:

Typing
Structured way of creating and interacting with object (even though you're using @ instread of $)

Overall 8/10

Reference Classes (This is where you clap hands)

Reference Classes:

node <- setRefClass("nodeR", fields = list(id = "numeric"), 
                     methods = list(
                       show = function(){
                          cat("Node\nID:",id,"\n")
                       }
                     ))
n1 <- node(id = 1)

print(n1)

Node
ID: 1

Benefits:

Unlike S3 and S4 methods belong to class
Able to have pointer like functionality
Powerful abstraction layer

Overall 10/10 !

plot of chunk unnamed-chunk-21

Reference Classes: Under the hood

Things to note:

Under the hood:
- A S4 class
- Its own environment

plot of chunk unnamed-chunk-22

Variables are references to underlying objects

Why worry?

Copying reference classes:

n1 <- node(id = 1)
n1

Node
ID: 1

Wrong way to copy:

n2 = n1
n2$id <- 2
n1

Node
ID: 2

plot of chunk unnamed-chunk-26

Reference Classes: Under the hood

Right way to copy:

n2 <- n1$copy()
n2$id <- 2
n1

Node
ID: 1

Let's get this party started!

Create some reference classes.

g<-dag()
g$load_from_file('classic_bst')
g$plot()

Other projections

g$plot_heirarchy()

Highlight features

g$plot_heirarchy(colorRootLeaves = T)

Highlight features

g$plot(colorRootLeaves = T)

Modify the structure

g$plot_heirarchy(colorRootLeaves = T, turn = T)

Modify the structure

g$close_graph()
g$plot_heirarchy(colorRootLeaves = T, turn = T)

Some graph search

We might be interested in a feature of the graph

shortest paths
critical paths
min/max cut
or other features

g$shortest_path(g$root_nodes(), g$leaf_nodes())

Some graph search

g$plot_heirarchy(T, T, g$shortest_path(g$root_nodes(), g$leaf_nodes()))

Some graph search

g$plot_heirarchy(T, T, g$critical_path(g$root_nodes(), g$leaf_nodes()))

Steady on

You're right, that was too quick. What just happened?

plot of chunk unnamed-chunk-37

Overkill?

No, just the right amount of kill.
Don't reinvent the wheel
Graphs can have multiple metrics/features
Extracting features consistently for search is troublesome
Which brings us to the Zebra suit.

Some graph search

Overengineering for the win

The stack:

Create reference class objects
- Nodes
- DAG
Implement algorithms using Rcpp
- Shortest path
- Path with highest awesomeness value
Visualise graphs using d3
BONUS ROUND: Create presentation using RPres
BONUS ROUND: Embed active charts

plot of chunk unnamed-chunk-38