Use case: Iterative Policy Evaluation (Reinforcement Learning)

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

In this vignette, we'll present a real-life use case, which shows how the matricks package makes the work with matrices easier.

Let's try to implement an algorithm from the field of Reinforcement Learning called iterative policy evaluation. The environment we will work on is a simple Grid World game. This vignette was inspired by a Reinforcement Learning course by Lazy Programmer, originally written in Python.

library(matricks)

Initializing environment

We initialize environment, using following functions from matrix package: * m() * matrix_idx() * plot.matrix() * seq_matrix() * sv() * with_same_dims()

We construct matrices, describing game environment: actions - all the possible moves rewards

# Defining possible actions
# FALSE = unavailable field
actions <- m(T, T, T, F, T|
             T, T, F, T, T|
             F, T, T, T, T|
             T, T, T, T, T|
             F, F, T, T, T|
             F, T, T, T, F)
plot(actions)
# Defining rewards matrix with two terminal states
rewards <- with_same_dims(actions, 0)
rewards <- sv(rewards,
              c(4, 4) ~ 1., # Win
              c(5, 3) ~ -1) # Lose
# We add small penalties, which encourages models
# to reduce number of steps it passes 
rewards[rewards == 0] <- -0.1
plot(rewards)
# List of possible state indices
states <- matrix_idx(actions, actions)

Setting fixed policy

Policy defines a set of moves we should perform, in order to achieve the best result in the game. We start with a fixed policy values.

# Symbols for moves
U <- "U" # Up
D <- "D" # Down
L <- "L" # Left
R <- "R" # Right

fixed.policy <- m(R , D , L , NA, D |
                  R , D , NA, D , L |
                  NA, R , D , D , D |
                  R , R , D , NA, D |
                  NA, NA, NA, U , L |
                  NA, R , U , L , NA)

It was easier to define policy using character symbols (as above). However, we have to transform into form that is more convenient for computations.

as_idx <- function(x){
  n.row <- nrow(x)
  n.col <- ncol(x)

  result <- matrix(list(), nrow = n.row, ncol = n.col)

  for (i in 1:n.row){
    for (j in 1:n.col){
      coords <- c(i, j)

      if (is.na(at(x, coords)))
        next

      x.val <- x[i, j]

      switch (x.val,
        U = c(i - 1, j),
        D = c(i + 1, j),
        L = c(i, j - 1),
        R = c(i, j + 1),
      ) -> move.idx

      if (!is_idx_possible(x, move.idx))
        next

      result[i, j] <- list(move.idx)
    }
  }
  result
}

fixed.policy.idx <- as_idx(fixed.policy)
fixed.policy.idx

Policy evaluation function

evaluate_policy <- function(actions, rewards, policy, 
                            epsilon = 1e-3, gamma = 0.9){

  V <- with_same_dims(policy, 0)
  policy.idx <- as_idx(policy)

  while (TRUE) {
    biggest.change <- 0

    for (move in seq_matrix(policy.idx)) {
      s <- move[[1]] # Action, value at index s
      a <- move[[2]]

      old.v <- at(V, s)

      if(!at(actions, s))
        next

      if(is.null(a))
        next

      r <- at(rewards, a)
      at(V, s) <- r + gamma * at(V, a)
      biggest.change <- max(biggest.change, abs(old.v - at(V, s)))
    }

    print(biggest.change)

    if (biggest.change < epsilon){
      break
    }
  }
  return(V)
}  

Experiment

V1 <- evaluate_policy(actions = actions, 
                      rewards = rewards,
                      policy  = fixed.policy)
V1
plot(V1)


Try the matricks package in your browser

Any scripts or data that you put into this service are public.

matricks documentation built on March 26, 2020, 6:22 p.m.