In Nth-iteration-labs/contextual: Simulation and Analysis of Contextual Multi-Armed Bandit Policies

knitr::opts_chunk$set(
  fig.pos = 'H',
  collapse = TRUE,
  comment = "#>"
)

Simulation of multi-armed Bandit policies following John Myles White's "Bandit algorithms for website_optimization".

The book, which offers a comprehensive entry-level introduction to context-free bandit policies, is available here: John Myles White. The orignal (Python) code for the book can be found in its complementary GitHub repository.

Discrepancies with the book

We did encounter one minor issue with the original code used to generate the plots in the book: its bandit policies always choose the first arm (in other words, the arm with the lowest index) when rewards between arms are tied. In the original Python code:

def ind_max(x):
  m = max(x)
  return x.index(m)

This introduces a bias that adds up over time, changing simulations' results and plots. In R, (which uses 1 based indexing) this corresponds to always doing:

rewards      <- c(0,0,1,0,1)
max_reward   <-which.max(rewards)
max_reward   # always contains lowest index, 3 in this case, biasing results

Contextual correctly returns the index of one of the tied maxima random, fixing the bias, resulting in slightly different plots. In Python, a patch along the following lines ought to resolve this issue:

def ind_max(x):
  max_value = max(x)
  max_keys = [k for k, v in enumerate(x) if v == max_value]
  return random.choice(max_keys)

In R, this corresponds to:

rewards      <- c(0,0,1,0,1)
max_reward   <- which(rewards == max(rewards))
if(length(max_reward) > 1){
  max_reward <- sample(max_reward, 1)
}
max_reward   # randomly assigned either 3 or 5, no bias

To illustrate, a replication of the original Figure 4-3 when running a simulation with the biased arm selection, (as used in the book) on the left, contextual's unbiased simulation to the right:

Chapter 4 - Debugging and epsilon greedy

prob_per_arm       <- c(0.1, 0.1, 0.1, 0.1, 0.9)
horizon            <- 250
simulations        <- 5000

bandit             <- BasicBernoulliBandit$new(prob_per_arm)

agents             <- list(Agent$new(EpsilonGreedyPolicy$new(0.1), bandit, "Epsilon = 0.1"),
                           Agent$new(EpsilonGreedyPolicy$new(0.2), bandit, "Epsilon = 0.2"),
                           Agent$new(EpsilonGreedyPolicy$new(0.3), bandit, "Epsilon = 0.3"),
                           Agent$new(EpsilonGreedyPolicy$new(0.4), bandit, "Epsilon = 0.4"),
                           Agent$new(EpsilonGreedyPolicy$new(0.5), bandit, "Epsilon = 0.5"))

simulation         <- Simulator$new(agents, horizon, simulations)
history            <- simulation$run()

Figure 4-2. How often does the epsilon greedy algorithm select the best arm?

plot(history, type = "optimal", legend_position = "bottomright", ylim = c(0,1))

Figure 4-3. How much reward does the epsilon greedy algorithm earn on average?

plot(history, type = "average", regret = FALSE, legend_position = "bottomright", ylim = c(0,1))

Figure 4-4. How much reward has the epsilon greedy algorithm earned by trial t?

plot(history, type = "cumulative", regret = FALSE)

Chapter 5 - Softmax

agents             <- list(Agent$new(SoftmaxPolicy$new(0.1), bandit, "Tau = 0.1"),
                           Agent$new(SoftmaxPolicy$new(0.2), bandit, "Tau = 0.2"),
                           Agent$new(SoftmaxPolicy$new(0.3), bandit, "Tau = 0.3"),
                           Agent$new(SoftmaxPolicy$new(0.4), bandit, "Tau = 0.4"),
                           Agent$new(SoftmaxPolicy$new(0.5), bandit, "Tau = 0.5"))

simulation         <- Simulator$new(agents, horizon, simulations)
history            <- simulation$run()

Figure 5-2. How often does the softmax algorithm select the best arm?

plot(history, type = "optimal", legend_position = "bottomright", ylim = c(0,1))

Figure 5-3. How much reward does the softmax algorithm earn on average?

plot(history, type = "average", regret = FALSE, legend_position = "bottomright", ylim = c(0,1))

Figure 5-4. How much reward has the softmax algorithm earned by trial t?

plot(history, type = "cumulative", regret = FALSE)

Chapter 6 - UCB

agents             <- list(Agent$new(SoftmaxPolicy$new(0.1), bandit, "Softmax"),
                           Agent$new(EpsilonGreedyPolicy$new(0.1), bandit, "EpsilonGreedy"),
                           Agent$new(UCB1Policy$new(), bandit, "UCB1"))

simulation         <- Simulator$new(agents, horizon, simulations)
history            <- simulation$run()

Figure 6-3. How often does the UCB algorithm select the best arm?

plot(history, type = "optimal", legend_position = "bottomright", ylim = c(0,1))

Figure 6-4. How much reward does the UCB algorithm earn on average?

plot(history, type = "average", regret = FALSE, legend_position = "bottomright", ylim = c(0,1))

Figure 6-5. How much reward has the UCB algorithm earned by trial t?

plot(history, type = "cumulative", regret = FALSE)

Nth-iteration-labs/contextual documentation built on July 28, 2020, 1:13 p.m.