/
bandits.R
204 lines (137 loc) · 4.67 KB
/
bandits.R
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
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
### Function for implementing Lakkaraju et al. (2017) bandits search algorithm
bandit_search <- function(D_test, c_MX, true_misclass, B, tau=.65, clust_max=5, clust_set=NULL, sigma=.001, scale=TRUE){
conf_cost <- function(c_MX,...) I(c_MX)
clust_out <- UUclust(D_test, c_MX, clust_max, clust_set) ## Create clusters+
k <- length((unique(clust_out)))
## Places to hold information
solution <- 0
utility <- c()
numFound <- 0
cumulativeConfidenceUnk <- 0
cumulativeConfidence <- 0
found <- 0
unscaledDist <- 1
unscaledReward <- 0
arms <- vector("list", k)
arm_rewards <- vector("list", k)
arm_times <- vector("list", k)
sim_mat <- make_sim_mat(D_test)
## Each arm is a cluster.
# Put the indicies of the instances from each cluster in the arm
arm_initial_size <- c()
for(i in sort(unique(clust_out))){
arms[i] = list(which(clust_out == i))
arm_initial_size <- c(arm_initial_size, length(arms[[i]]))
}
## Helper functions
disc <- function(arm, j, t){
partOne <- arm_initial_size[arm] - length(arm_times[[arm]])
partTwo <- arm_initial_size[arm] - length(which(arm_times[[arm]] <= j))
return(partOne/partTwo)
}
n <- function(arm){
return(length(arm_times[[arm]]))
}
u <- function(arm){
return(sum(arm_rewards[[arm]])/n(arm))
}
n_disc <- function(arm, t){
toReturn <- 0
for(j in arm_times[[arm]]){
toReturn <- toReturn + disc(arm, j, t)
}
return(toReturn)
}
## Count number of uknowns found in that arm up to time t
u_disc <- function(arm, t){
toReturn <- 0
for(i in 1:length(arm_times[[arm]])){
partOne <- disc(arm, arm_times[[arm]][i], t) * arm_rewards[[arm]][i]
toReturn <- toReturn + partOne
}
return(toReturn/n_disc(arm, t))
}
isvalid <- function(arm){
if(length(arms[[arm]]) > 0){
return(1)
}
return(0)
}
## List of valid arms
#
ne_arms <- c()
for(i in 1:k){
if(isvalid(i)){
ne_arms <- c(ne_arms, i)
}
}
## Do this until out of budget
for(t in 1:B){
## Find current arm
# If not all arms sampled, curr_arm is the next arm
# Otherwise curr_arm is the most promising arm
curr_arm <- -10
if(t <= length(ne_arms)){
curr_arm <- ne_arms[t]
}else{
findMax <- c()
for(i in 1:k){
if(isvalid(i) == 1){
partOne <- u_disc(i, t)
partTwo <- sqrt(2*log(t)/n_disc(i, t))
findMax <- c(findMax, partOne + partTwo)
}else{
findMax <- c(findMax, 0)
}
}
curr_arm <- which.max(findMax)
}
## Find instance in arm to select
toAdd <- 0
if(length(arms[[curr_arm]]) != 1){
# find optimal new q observations
Sc_idx <- arms[[curr_arm]]
P_expx_Q <- P_explainx_Q(sim_mat, solution, true_misclass, c_MX, tau)
q_new_idx <- Sc_idx[which.max(exp_utility_step_all(Sc_idx, rep(1/length(c_MX), length(P_expx_Q)), sim_mat, c_MX, P_expx_Q, "conf_cost"))]
toAdd <- q_new_idx
}else{
toAdd <- arms[[curr_arm]][1]
}
## Remove selected instance from arm
arms[[curr_arm]] = arms[[curr_arm]][which(arms[[curr_arm]] != toAdd)]
## Add instance to solution
solution[t] <- toAdd
## Update utility
P_expx_Q <- P_explainx_Q(sim_mat, solution, true_misclass, c_MX, tau)
utility <- c(utility, t(get("conf_cost")(c_MX, rep(1/length(c_MX), length(P_expx_Q))) %*% P_expx_Q))
## Track reward
reward <- utility[length(utility)]
if(length(utility) > 1){
reward <- utility[length(utility)] - utility[length(utility)-1]
}
if(reward > 0.01){
reward = 1
}else{
reward = 0
}
arm_rewards[[curr_arm]] <- c(arm_rewards[[curr_arm]], reward)
arm_times[[curr_arm]] <- c(arm_times[[curr_arm]], t)
## Track a bunch of other stuff
S_idx <- solution[true_misclass[solution]==1]
cumulativeConfidence[t] <- ifelse(is.null(solution), 0, sum(c_MX[solution]))
cumulativeConfidenceUnk[t] <- ifelse(is.null(S_idx), 0, sum(c_MX[S_idx]))
unscaledDist[t] <- sum(min_dist(sim_mat, S_idx))/length(c_MX)
unscaledReward[t] <- ifelse(is.null(S_idx), 0, sum(cost_c_MX(c_MX[S_idx])))
found[t] <- length(S_idx)
}
return(data.frame(cost="cost_conf",
phi="Bandits",
utility=utility,
Q_idx=solution,
b=1:B,
cumulativeConfidence = cumulativeConfidence,
cumulativeConfidenceUnk = cumulativeConfidenceUnk,
unscaledDist = unscaledDist,
unscaledReward = unscaledReward,
found = found))
}