Wednesday, December 18, 2013

Reinforcement Learning in Scala: Model

Target audience: Advanced
Estimated reading time: 5'

This post is the second installment of our introduction to reinforcement learning using the temporal difference. This section is dedicated to the ubiquitous Q-learning algorithm.

Table of contents
Follow me on LinkedIn

Overview

The previous post, Reinforcement learning in Scala: Policies introduced the concept of reinforcement learning, temporal difference and more specifically the Q-learning algorithm:
Reinforcement Learning I: States & Policy The following components were implemented:
  • States QLState and their associated actions QLAction
  • States space QLSpace
  • Policy QLPolicy as a container of tuple {reward, Q-value, probabilities} of the Q-learning model
The last two components to complete the implementation of Q-learning are the model and the training algorithm.

Modeling and training

The first step is to define a model for the reinforcement learning. A model is created during training and is composed of
  • Best policy to transition from any initial state to a goal state
  • Coverage ratio as defined as the percentage of training cycles that reach the (or one of the) goal.
class QLModel[T](val bestPolicy: QLPolicy[T], val coverage: Double)


The QLearning class takes 3 arguments
  • A set of configuration parameters config
  • The search/states space qlSpace
  • The initial policy associated with the states (reward and probabilities) qlPolicy

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
class QLearning[T](
   config: QLConfig, 
   qlSpace: QLSpace[T], 
   qlPolicy: QLPolicy[T]) 

    //model in Q-learning algorithm
  val model: Option[QLModel[T]] = train.toOption
    
    // Generate a model through multi-epoch training
  def train: Try[Option[QLModel[T]]] {}
  private def train(r: Random): Boolean {}

   // Predict a state as a destination of this current 
   // state, given a model
  def predict : PartialFunction[QLState[T], QLState[T]] {}

  // Select next state and action index
  def nextState(st: (QLState[T], Int)): (QLState[T], Int) {} 
}

The model of type Option[QLModel] (line 7) is created by the method train (line 10). Its value is None if training failed.

The training method train consists of executing config.numEpisodes cycle or episode of a sequence of state transition (line 5). The random generator r is used in the initialization of the search space.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
def train: Option[QLModel[T]] = {
  val r = new Random(System.currentTimeMillis)

  Try {
    val completions = Range(0, config.numEpisodes).filter(train(r) )

    val coverage = completions.toSize.toDouble/config.numEpisodes
    if(coverage > config.minCoverage) 
       new QLModel[T](qlPolicy, coverage)
    else 
       QLModel.empty[T]
  }.toOption
}

The training process exits with the model if the minimum minCoverage (number of episodes for which the goal state is reached) is met (line 8).

The method train(r: scala.util.Random) uses a tail recursion to transition from the initial random state to one of the goal state. The tail recursion is implemented by the search method (line 4). The method implements the recursive temporal difference formula introduced in the previous post (Reinforcement Learning I: States & Policy) (lines 14-18). 

The state for which the action generates the highest reward R given a policy qlPolicy (line 10) is computed for each new state transition. The Q-value of the current policy is then updated qlPolicy.setQ before repeating the process for the next state, through recursion (line 21).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
def train(r: Random): Boolean = {
   
  @scala.annotation.tailrec
  def search(st: (QLState[T], Int)): (QLState[T], Int) = {
    val states = qlSpace.nextStates(st._1)
    if( states.isEmpty || st._2 >= config.episodeLength ) 
        (st._1, -1)
    
    else {
      val state = states.maxBy(s => qlPolicy.R(st._1.id, s.id))
      if( qlSpace.isGoal(state) )
          (state, st._2)
      else {
        val r = qlPolicy.R(st._1.id, state.id)   
        val q = qlPolicy.Q(st._1.id, state.id)
        // Q-Learning formula
        val deltaQ = r + config.gamma*qlSpace.maxQ(state, qlPolicy) -q
        val nq = q + config.alpha*deltaQ
        
        qlPolicy.setQ(st._1.id, state.id,  nq)
        search((state, st._2+1))
       }
     }
  } 
   
  r.setSeed(System.currentTimeMillis*Random.nextInt)

  val finalState = search((qlSpace.init(r), 0))
  if( finalState._2 != -1) 
     qlSpace.isGoal(finalState._1) 
  else 
     false
}

Note: There is no guaranty that one of the goal state is reached from any initial state chosen randomly. It is expected that some of the training epoch fails. This is the reason why monitoring coverage is critical. Obviously, you may choose a deterministic approach to the initialization of each training epoch by picking up any state beside the goal state(s), as a starting state.

Prediction

Once trained, the model is used to predict the next state with the highest value (or probability) given an existing state. The prediction is implemented as a partial function.

1
2
3
4
def predict : PartialFunction[QLState[T], QLState[T]] = {
  case state: QLState[T] if(model != None) => 
    if( state.isGoal) state else nextState(state, 0)._1
}

The method nextState does the heavy lifting. It retrieves the list of states associated with the current state st through its actions set (line 2). The next most rewarding state qState is computed using the reward matrix R of the best policy of the QLearning model (lines 6 - 8).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
def nextState(st: (QLState[T], Int)): (QLState[T], Int) =  {
  val states = qlSpace.nextStates(st._1)
  if( states.isEmpty || st._2 >= config.episodeLength) 
    st
  else {
    val qState = states.maxBy(
     s => model.map(_.bestPolicy.R(st._1.id, s.id))
           .getOrElse(-1.0)
    )

    nextState( (qState, st._2+1))
  }
}


References

Tuesday, December 10, 2013

Reinforcement Learning in Scala: States & Policies

Target audience: Advanced
Estimated reading time: 9'

This post describes a very common reinforcement learning methodology: Tenporal difference update as implemented in Scala. This first section introduces and implements the concept of states and policies.


Table of contents
Follow me on LinkedIn

Overview

There are many different approaches to implement reinforcement learning
One of the most commonly used method is searching the value function space using temporal difference method
All known reinforcement learning methods share the same objective of solving the sequential decision tasks. In a sequential decision task, an agent interacts with a dynamic system by selecting actions that affect the transition between states in order to optimize a given reward function.


At any given step i, the agent select an action a(i) on the current state s(i). The dynamic system responds by rewarding the agent for its optimal selection of the next state:\[s_{i+1}=V(s_{i})\]
The learning agent infers the policy that maps the set of states {s} to the set of available actions {a}, using a value function  \[V(s_{i})\] The policy is defined at \[\pi :\,\{s_{i}\} \mapsto \{a_{i}\} \left \{ s_{i}|s_{i+1}=V(s_{i}) \right \}\]


Temporal difference

The most common approach of learning a value function V is to use the Temporal Difference method (TD). The method uses observations of prediction differences from consecutive states, s(i) & s(i+1). If we note r the reward for selection an action from state s(i) to s(i+1) and n the learning rate, then the value V is updated as \[V(s_{i})\leftarrow V(s_{i})+\eta .(V(s_{i+1}) -V(s_{i}) + r_{i})\]
Therefore the goal of the temporal difference method is to learn the value function for the optimal policy. The 'action-value' function represents the expected value of action a on a state s and defined as \[Q(s_{i},a_{i}) = r(s_{i}) + V(s_{i})\] where r is the reward value for the state.


On-policy vs. off-policy

The Temporal Difference method relies on the estimate of the final reward to be computed for each state. There are two methods of the Temporal Difference algorithm:On-Policy and Off-Policy:
  - On-Policy method learns the value of the policy used to make the decision. The value function is derived from the execution of actions using the same policy but based on history
 - Off-Policy method learns potentially different policies. Therefore the estimate is computed using actions that have not been executed yet.

The most common formula for temporal difference approach is the Q-learning formula. It introduces the concept of discount rate to reduce the impact of the first few states on the optimization of the policy. It does not need a model of its environment. The exploitation of action-value approach consists of selecting the next state is by computing the action with the maximum reward. Conversely the exploration approach focus on the total anticipated reward.The update equation for the Q-Learning is \[Q(s_{i},a_{i}) \leftarrow Q(s_{i},a_{i}) + \eta .(r_{i+1} +\alpha .max_{a_{i+1}}Q(s_{i+1},a_{i+1}) - Q(s_{i},a_{i}))\] \[Q(s_{i},a_{i}): \mathrm{expected\,value\,action\,a\,on\,state\,s}\,\,\eta : \mathrm{learning\,rate}\,\,\alpha : \mathrm{discount\,rate}\] . One of the most commonly used On-Policy method is Sarsa which does not necessarily select the action that offer the most value.The update equation is defined as\[Q(s_{i},a_{i}) \leftarrow Q(s_{i},a_{i}) + \eta .(r_{i+1} +\alpha .Q(s_{i+1},a_{i+1}) - Q(s_{i},a_{i}))\]

States and actions

Functional languages are particularly suitable for iterative computation. We use Scala for the implementation of the temporal difference algorithm. We allow the user to specify any variant of the learning formula, using local functions or closures.
Firstly, we have to define a state class, QLState (line 1) that contains a list of actions of type QLAction (line 3) that can be executed from this state. The only purpose of this class is to connect a list of action to a source state. The parameterized class argument property (line 4) is used to "attach" some extra characteristics to this state. 

1
2
3
4
5
6
7
8
class QLState[T](
  val id: Int, 
  val actions: List[QLAction[T]] = List.empty, 
  property: T) {
    
  @inline
  def isGoal: Boolean = !actions.isEmpty
}

As described in the introduction, an action of class QLAction has a source state from and a destination state to(state which is reached following the action). A state except the goal state, has multiple actions but an action has only one destination or resulting state.

case class QLAction[T](from: Int, to: Int)


The state and action can be loaded, generated and managed by a directed graph or search space of type QLSpace. The search space contains the list of all the possible states available to the agent.
One or more of these states can be selected as goals. The algorithm does not restrict the agent to a single state. The process ends when one of the goal states is reached (OR logic). The algorithm does not support combined goals (AND logic).

Let's implement the basic components of the search space QLSpace. The class list all available states (line 2) and one or more final or goal states goalIds (line 3). Although you would expect that the search space contains a single final or goal state, it is not uncommon to have online training using more than one goal state.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
class QLSpace[T](
   states: Array[QLState[T]], 
   goalIds: Array[Int]) {

    // Indexed map of states 
  val statesMap: immutable.Map[Int, QLState[T]] = 
    states.map(st => (st.id, st)).toMap
    // List set of one or more goals  
  val goalStates = new immutable.HashSet[Int]() ++ goalIds
 
    // Compute the maximum Q value for a given state and policy
  def maxQ(st: QLState[T], policy: QLPolicy[T]): Double = { 
    val best = states.filter( _ != st)
       .maxBy(_st => policy.EQ(st.id, _st.id))
    policy.EQ(st.id, best.id)
  }
 
    // Retrieves the list of states destination of state, st
  def nextStates(st: QLState[T]): List[QLState[T]] =
     st.actions.map(ac => statesMap.get(ac.to).get)
 
  def init(r: Random): QLState[T] = 
    states(r.nextInt(states.size-1))
}

A hash map statesMap maintains a dictionary of all the possible states with the state id as unique key (lines 6, 7). The class QLSpace has three important methods:
  • init initializes the search with a random state for each training epoch (lines 22, 23)
  • nextStates returns the list of destination states associated to the state st (lines 19, 20)
  • maxQ return the maximum Q-value for this state st given the current policy policy(lines 12-15). The method filters out itself from the search from the next best action. It then compute the maximum reward or Q(state, action) value according to the given policy policy
The next step is to defined a policy.

Learning policy

A policy is defined by three components
  • A reward collected after transitioning from one state to another state (line 2). The reward is provided by the user
  • A Q(State, Action) value, value associated to a transition state and an action (line 4)
  • A probability (with default values of 1.0) that defines the obstacles or hindrance to migrate from one state to another (line 3)
The estimate combine the Q-value (incentive to move to the best next step) and probability (hindrance to move to any particular state) (line 7).

1
2
3
4
5
6
7
8
class QLData {
  var reward: Double = 1.0
  var probability: Double = 1.0
  var value: Double = 0.0) {
  
  @inline
  final def estimate: Double = value*probability
}

The policy of type QLPolicy is a container for the state transition attributes, rewards, Q-values and probabilities.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
class QLPolicy[T](numStates: Int, input: Array[QLInput]) {
 
  val qlData = {
    val data = Array.tabulate(numStates)(
      _ => Array.fill(numStates)(new QLData)
    )
 
    input.foreach(in => {  
      data(in.from)(in.to).reward = in.reward
      data(in.from)(in.to).probability = in.prob
    })
    data
  }
  
  def setQ(from: Int, to: Int, value: Double): Unit =
     qlData(from)(to).value = value
 
  def Q(from: Int, to: Int): Double = qlData(from)(to).value
}

The constructor for QLPolicy takes two arguments:
  • Number of states numStates (line 1)
  • Sequence of input of type QLInput to the policy
The constructor create a numStates x numStates matrix of transition of type QLData (lines 3 - 12), from the input.

The type QLInput wraps the input data (index of the input state from, index of the output state to, reward and probability associated to the state transition) into a single convenient class
.

case class QLInput(
   from: Int, 
   to: Int, 
   reward: Double = 1.0, 
   prob: Double = 1.0
)

The next post will dig into the generation of a model through Q-learning training


References

Sunday, November 3, 2013

Breakable Loops in Scala

Target audience: Beginner
Estimated reading time: 4'



Table of contents
Follow me on LinkedIn

Introduction

Contrary to C++ and Java, Scala does not allow developer to exit an iterative execution prematurely using a syntax equivalent to break.

var sum = 0

for( i <- 0 until 100) {
  sum += i
  if( sum > 400) 
    break   // won't compile!                                  
}

Scala purists tend to stay away from this type of constructs and use higher order collection methods such as
  exists( p: (T) => Boolean)
   find( p: (T) => Boolean)
takeWhile( p: (T) => Boolean
)
However these methods are not available outside collections. There are cases where a `break` construct may be a simpler solution.
This post review the different options to 'break' from a loop according to a predefined condition.

Breakable control

Although breaking out of a loop may not be appealing to "pure" functional developer, the Scala language provides formal Java and C++ developers with the option to use break and continue statements. The breakable statement define the scope for which break is valid.

import scala.util.control.Breaks._                            

var sum = 0
breakable { 
  for (i <- 0 until 100 ) {
    sum += i
    if( sum > 55) 
      break
  }
}

Any experienced developer would be quite reluctant to use such an idiom: beside the fact that the accumulator is defined as a variable, the implementation is unnecessary convoluted adding a wrapper around the loop. Let's try to find a more functional like approach to breaking out of any loop.

scan & fold to the rescue

Luckily, Scala provides collections with functional traversal patterns that can be used and chained to break, elegantly from a loop. The following code snippets introduce applies some of those traversal patterns to an array and a associative map to illustrate the overall "functional" approach to iteration. 
Let's consider the problem of extracting the elements of an array or map before a predefined condition is met for an element. For instance, let's extract the elements of an array or a hash map until one element has the value 0. The Scala programming language provide us with the takeWhile method (lines 7 & 10) that allows to to end traversing a collection and return a elements of the collection visited when a condition is met.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
val values = Array[Int](9, 45, 11, 0, 67, 33)

val mappedValues = Map[String, Int](
  "A"->2, "B"->45, "C"->67, "D"->0, "E"->56
)
      // Extract a subarray of values until the condition is met
values takeWhile( _ != 0) // -> Array(9, 45, 11)
 
     // Extract a submap until the condition is met
mappedValues takeWhile( _._2 != 0) 
     // -> Map(E -> 56, A -> 2, B -> 45, C -> 67)

The second case consists in accumulating values until a condition on the accumulator is met. Contrary to fold and reduce methods which apply a function (i.e summation) to all the elements of a collection to return a single value, scan, scanLeft and scanRight methods return a collection of values processed by the function.

In the following example, the invocation of the higher order method scanLeft (lines 4 generates an array and a hash map with the cumulative values. The takeWhile method i(lines 5 & 8) s then applied to the resulting array or map to return a collection of cumulative values until the accumulator exceeds 56. The same invocation can be used to return the element which push the accumulator beyond the threshold.

1
2
3
4
5
6
7
8
9
values.scanLeft(0)(_ + _).takeWhile (_ < 56) 
   //-> Array(0, 9, 54)
 
mappedValues.scanLeft(0) (_ + _._2)
   .takeWhile( _ < 56)
 
val indexTargetEl = values.scanLeft(0)(_ + _)
      .takeWhile (_ < 56).size -1            
val targetEl = values(indexTargetEl)



Tail recursion

A third and elegant alternative to exit from a loop is using the tail recursion which is supported natively in the Scala language. The first method, findValue exits the loop when a condition on the value is reached (line 5). The second method, findCummulative, is implemented as a closure and exits when the sum of elements exceeds a predefined value (line 15).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
val values = Array[Int](9, 45, 11, 0, 67, 33)
 
@scala.annotation.tailrec
def findValue(values: Array[Int], index: Int) : Int = {
   if( values(index) == 0 || index >= values.size)
     index
   else
     findValue(values, index + 1)
}
val newValues = values.slice(0, findValue(values, 0))   

val MAX_VALUE :Int = 86

@scala.annotation.tailrec
def findCumulative(sum: Int, index: Int): Int = {
   if( index >= values.size ||sum >= MAX_VALUE)
     index -1
   else
     findCumulative(sum + values(index), index + 1)
}

val newValues2 = values.slice(0, findCumulative(0, 0))

This concludes our review of three different Scala constructs available to developer to exit prematurely from a loop
  • breakable construct
  • Higher order method such as exists, find, takeWhile....
  • tail recursion

References