我正在使用 Sutton & Barto 的电子书强化学习：强化学习研究简介。我在尝试模拟action-value page上的结果（图）时遇到了一些问题。
更具体地说，如何模拟greedy每个任务的价值？书上说：
...我们可以绘制各种方法的性能和行为，因为它们随着超过 1000 次播放的经验而改进...
所以我想我必须跟踪探索性价值，因为找到更好的价值。问题是如何使用贪婪方法来做到这一点 - 因为没有探索性的举动，我怎么知道什么是贪婪行为？
感谢所有的评论和回答！
更新：请参阅我的答案中的代码。
我终于做对了。正如书中指出的那样，eps 玩家应该因为探索性动作而击败贪婪的玩家。代码很慢，需要一些优化，但这里是：
get.testbed = function(arms = 10, plays = 500, u = 0, sdev.arm = 1, sdev.rewards = 1){
optimal = rnorm(arms, u, sdev.arm)
rewards = sapply(optimal, function(x)rnorm(plays, x, sdev.rewards))
list(optimal = optimal, rewards = rewards)
}
play.slots = function(arms = 10, plays = 500, u = 0, sdev.arm = 1, sdev.rewards = 1, eps = 0.1){
testbed = get.testbed(arms, plays, u, sdev.arm, sdev.rewards)
optimal = testbed$optimal
rewards = testbed$rewards
optim.index = which.max(optimal)
slot.rewards = rep(0, arms)
reward.hist = rep(0, plays)
optimal.hist = rep(0, plays)
pulls = rep(0, arms)
probs = runif(plays)
# vetorizar
for (i in 1:plays){
## dont use ifelse() in this case
## idx = ifelse(probs[i] < eps, sample(arms, 1), which.max(slot.rewards))
idx = if (probs[i] < eps) sample(arms, 1) else which.max(slot.rewards)
reward.hist[i] = rewards[i, idx]
if (idx == optim.index)
optimal.hist[i] = 1
slot.rewards[idx] = slot.rewards[idx] + (rewards[i, idx] - slot.rewards[idx])/(pulls[idx] + 1)
pulls[idx] = pulls[idx] + 1
}
list(slot.rewards = slot.rewards, reward.hist = reward.hist, optimal.hist = optimal.hist, pulls = pulls)
}
do.simulation = function(N = 100, arms = 10, plays = 500, u = 0, sdev.arm = 1, sdev.rewards = 1, eps = c(0.0, 0.01, 0.1)){
n.players = length(eps)
col.names = paste('eps', eps)
rewards.hist = matrix(0, nrow = plays, ncol = n.players)
optim.hist = matrix(0, nrow = plays, ncol = n.players)
colnames(rewards.hist) = col.names
colnames(optim.hist) = col.names
for (p in 1:n.players){
for (i in 1:N){
play.results = play.slots(arms, plays, u, sdev.arm, sdev.rewards, eps[p])
rewards.hist[, p] = rewards.hist[, p] + play.results$reward.hist
optim.hist[, p] = optim.hist[, p] + play.results$optimal.hist
}
}
rewards.hist = rewards.hist/N
optim.hist = optim.hist/N
optim.hist = apply(optim.hist, 2, function(x)cumsum(x)/(1:plays))
### Plot helper ###
plot.result = function(x, n.series, colors, leg.names, ...){
for (i in 1:n.series){
if (i == 1)
plot.ts(x[, i], ylim = 2*range(x), col = colors[i], ...)
else
lines(x[, i], col = colors[i], ...)
grid(col = 'lightgray')
}
legend('topleft', leg.names, col = colors, lwd = 2, cex = 0.6, box.lwd = NA)
}
### Plot helper ###
#### Plots ####
require(RColorBrewer)
colors = brewer.pal(n.players + 3, 'Set2')
op <-par(mfrow = c(2, 1), no.readonly = TRUE)
plot.result(rewards.hist, n.players, colors, col.names, xlab = 'Plays', ylab = 'Average reward', lwd = 2)
plot.result(optim.hist, n.players, colors, col.names, xlab = 'Plays', ylab = 'Optimal move %', lwd = 2)
#### Plots ####
par(op)
}
要运行它，只需调用
do.simulation(N = 100, arms = 10, eps = c(0, 0.01, 0.1))
您还可以选择使用 R 包“ contextual ”，它旨在简化上下文无关（如 Sutton 和 Barto 中所述）和上下文（例如 LinUCB）多臂强盗策略的实施和评估.
该软件包实际上提供了一个关于如何复制所有 Sutton 和 Barto 强盗图的小插曲。例如，要生成 ε-greedy 图，只需针对高斯老虎机模拟EpsilonGreedy策略：
library(contextual)
set.seed(2)
mus <- rnorm(10, 0, 1)
sigmas <- rep(1, 10)
bandit <- BasicGaussianBandit$new(mu_per_arm = mus, sigma_per_arm = sigmas)
agents <- list(Agent$new(EpsilonGreedyPolicy$new(0), bandit, "e = 0, greedy"),
Agent$new(EpsilonGreedyPolicy$new(0.1), bandit, "e = 0.1"),
Agent$new(EpsilonGreedyPolicy$new(0.01), bandit, "e = 0.01"))
simulator <- Simulator$new(agents = agents, horizon = 1000, simulations = 2000)
history <- simulator$run()
plot(history, type = "average", regret = FALSE, lwd = 1, legend_position = "bottomright")
plot(history, type = "optimal", lwd = 1, legend_position = "bottomright")
全面披露：我是该软件包的开发人员之一。
根据我们的聊天，这是我迄今为止所拥有的：
set.seed(1)
getRewardsGaussian <- function(arms, plays) {
## assuming each action has a normal distribution
# first generate new means
QStar <- rnorm(arms, 0, 1)
# then for each mean, generate `play`-many samples
sapply(QStar, function(u)
rnorm(plays, u, 1))
}
CalculateRewardsPerMethod <- function(arms=7, epsi1=0.01, epsi2=0.1
, plays=1000, methods=c("greedy", "epsi1", "epsi2")) {
# names for easy handling
names(methods) <- methods
arm.names <- paste0("Arm", ifelse((1:arms)<10, 0, ""), 1:arms)
# this could be different if not all actions' rewards have a gaussian dist.
rewards.source <- getRewardsGaussian(arms, plays)
# Three dimensional array to track running averages of each method
running.avgs <-
array(0, dim=c(plays, arms, length(methods))
, dimnames=list(PlayNo.=NULL, Arm=arm.names, Method=methods))
# Three dimensional array to track the outcome of each play, according to each method
rewards.received <-
array(NA_real_, dim=c(plays, 2, length(methods))
, dimnames=list(PlayNo.=seq(plays), Outcome=c("Arm", "Reward"), Method=methods))
# define the function internally to not have to pass running.avgs
chooseAnArm <- function(p) {
# Note that in a tie, which.max returns the lowest value, which is what we want
maxes <- apply(running.avgs[p, ,methods, drop=FALSE], 3, which.max)
# Note: deliberately drawing two separate random numbers and keeping this as
# two lines of code to accent that the two draws should not be related
if(runif(1) < epsi1)
maxes["epsi1"] <- sample(arms, 1)
if(runif(1) < epsi2)
maxes["epsi2"] <- sample(arms, 1)
return(maxes)
}
## TODO: Perform each action at least once, then select according to algorithm
## Starting points. Everyone starts at machine 3
choice <- c(3, 3, 3)
reward <- rewards.source[1, choice]
## First run, slightly different
rewards.received[1,,] <- rbind(choice, reward)
running.avgs[1, choice, ] <- reward # if different starting points, this needs to change like below
## HERE IS WHERE WE START PULLING THE LEVERS ##
## ----------------------------------------- ##
for (p in 2:plays) {
choice <- chooseAnArm(p)
reward <- rewards.source[p, choice]
# Note: When dropping a dim, the methods will be the columns
# and the Outcome info will be the rows. Use `rbind` instead of `cbind`.
rewards.received[p,,names(choice)] <- rbind(choice, reward)
## Update the running averages.
## For each method, the current running averages are the same as the
## previous for all arms, except for the one chosen this round.
## Thus start with last round's averages, then update the one arm.
running.avgs[p,,] <- running.avgs[p-1,,]
# The updating is only involved part (due to lots of array-indexing)
running.avgs[p,,][cbind(choice, 1:3)] <-
sapply(names(choice), function(m)
# Update the running average for the selected arm (for the current play & method)
mean( rewards.received[ 1:p,,,drop=FALSE][ rewards.received[1:p,"Arm",m] == choice[m],"Reward",m])
)
} # end for-loop
## DIFFERENT RETURN OPTIONS ##
## ------------------------ ##
## All rewards received, in simplifed matrix (dropping information on arm chosen)
# return(rewards.received[, "Reward", ])
## All rewards received, along with which arm chosen:
# return(rewards.received)
## Running averages of the rewards received by method
return( apply(rewards.received[, "Reward", ], 2, cumsum) / (1:plays) )
}
### EXECUTION (AND SIMULATION)
## PARAMETERS
arms <- 10
plays <- 1000
epsi1 <- 0.01
epsi2 <- 0.1
simuls <- 50 # 2000
methods=c("greedy", "epsi1", "epsi2")
## Single Iteration:
### we can run system time to get an idea for how long one will take
tme <- system.time( CalculateRewardsPerMethod(arms=arms, epsi1=epsi1, epsi2=epsi2, plays=plays) )
cat("Expected run time is approx: ", round((simuls * tme[["elapsed"]]) / 60, 1), " minutes")
## Multiple iterations (simulations)
rewards.received.list <- replicate(simuls, CalculateRewardsPerMethod(arms=arms, epsi1=epsi1, epsi2=epsi2, plays=plays), simplify="array")
## Compute average across simulations
rewards.received <- apply(rewards.received.list, 1:2, mean)
## RESULTS
head(rewards.received, 17)
MeanRewards <- rewards.received
## If using an alternate return method in `Calculate..` use the two lines below to calculate running avg
# CumulRewards <- apply(rewards.received, 2, cumsum)
# MeanRewards <- CumulRewards / (1:plays)
## PLOT
plot.ts(MeanRewards[, "greedy"], col = 'red', lwd = 2, ylim = range(MeanRewards), ylab = 'Average reward', xlab="Plays")
lines(MeanRewards[, "epsi1"], col = 'orange', lwd = 2)
lines(MeanRewards[, "epsi2"], col = 'navy', lwd = 2)
grid(col = 'darkgray')
legend('bottomright', c('greedy', paste("epsi1 =", epsi1), paste("epsi2 =", epsi2)), col = c('red', 'orange', 'navy'), lwd = 2, cex = 0.8)
您可能还想查看此链接 https://www.datahubbs.com/multi_armed_bandits_reinforcement_learning_1/
来自上述来源的相关代码的副本它不使用 R，而只是来自 numpy 的 np.random.rand()
class eps_bandit:
'''
epsilon-greedy k-bandit problem
Inputs
=====================================================
k: number of arms (int)
eps: probability of random action 0 < eps < 1 (float)
iters: number of steps (int)
mu: set the average rewards for each of the k-arms.
Set to "random" for the rewards to be selected from
a normal distribution with mean = 0.
Set to "sequence" for the means to be ordered from
0 to k-1.
Pass a list or array of length = k for user-defined
values.
'''
def __init__(self, k, eps, iters, mu='random'):
# Number of arms
self.k = k
# Search probability
self.eps = eps
# Number of iterations
self.iters = iters
# Step count
self.n = 0
# Step count for each arm
self.k_n = np.zeros(k)
# Total mean reward
self.mean_reward = 0
self.reward = np.zeros(iters)
# Mean reward for each arm
self.k_reward = np.zeros(k)
if type(mu) == list or type(mu).__module__ == np.__name__:
# User-defined averages
self.mu = np.array(mu)
elif mu == 'random':
# Draw means from probability distribution
self.mu = np.random.normal(0, 1, k)
elif mu == 'sequence':
# Increase the mean for each arm by one
self.mu = np.linspace(0, k-1, k)
def pull(self):
# Generate random number
p = np.random.rand()
if self.eps == 0 and self.n == 0:
a = np.random.choice(self.k)
elif p < self.eps:
# Randomly select an action
a = np.random.choice(self.k)
else:
# Take greedy action
a = np.argmax(self.k_reward)
reward = np.random.normal(self.mu[a], 1)
# Update counts
self.n += 1
self.k_n[a] += 1
# Update total
self.mean_reward = self.mean_reward + (
reward - self.mean_reward) / self.n
# Update results for a_k
self.k_reward[a] = self.k_reward[a] + (
reward - self.k_reward[a]) / self.k_n[a]
def run(self):
for i in range(self.iters):
self.pull()
self.reward[i] = self.mean_reward
def reset(self):
# Resets results while keeping settings
self.n = 0
self.k_n = np.zeros(k)
self.mean_reward = 0
self.reward = np.zeros(iters)
self.k_reward = np.zeros(k)