-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmain.R
197 lines (159 loc) · 7.48 KB
/
main.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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
require(pso)
#source("fit.R")
source("graphics.R")
order_dir = "orderings/"
model_dir = "models/"
data_dir = "data/"
# load lists of many trial orderings (and some means of human performance):
load(paste(order_dir,"master_orders.RData",sep='')) # orders
load(paste(order_dir,"asymmetric_conditions.RData",sep='')) # conds
print(names(orders))
print(names(conds))
load(paste(data_dir,"asym_master.RData",sep='')) # raw
run_model <- function(cond, model_name, parameters, print_perf=F) {
require(pso) # or require(DEoptim)
source(paste(model_dir,model_name,".R",sep=''))
mod = model(parameters, ord=cond$train)
if(print_perf) print(mod$perf)
return(mod)
}
animate_trajectories <- function(models, conds, condnames, params) {
for(m in models) {
for(cname in condnames) {
mod = run_model(conds[[cname]], m, params, print_perf=T)
animate_trajectory(mod, m, cname)
}
}
}
models = c("kachergis", "strength", "uncertainty", "novelty", "Bayesian_decay", "rescorla_wagner")
condnames = c("orig_3x3","freq369-3x3hiCD","freq369-3x3loCD")
models = c("kachergis","fazly","rescorla-wagner")
params = list("kachergis"=c(0.02470323,0.4890386,0.977897),
"fazly"=c(0.01557498,20000,0.1003937),
"rescorla-wagner"=c(0.08853654,0.9636937,6.48722)) # best from cond 207 (arbitrarily) 4x4 + 2 w/o
animate_trajectories(models, orders, condnames, params)
#mod = run_model(conds[["201"]], "fazly", c(.0001,8000,.7), print_perf=T)
#animate_trajectory(mod, "fazly", "201")
mod = run_model(conds[["201"]], "kachergis", c(1,3,.97), print_perf=T)
mod = run_model(conds[["201"]], "strength", c(1,.97), print_perf=T)
mod = run_model(conds[["201"]], "uncertainty", c(1,3,.97), print_perf=T)
mod = run_model(conds[["201"]], "novelty", c(1,3,.97), print_perf=T)
mod = run_model(conds[["201"]], "Bayesian_decay", c(.7,1.7,1), print_perf=T)
mod = run_model(conds[["201"]], "rescorla-wagner", c(.7,1.7,1), print_perf=T)
coocs3x4 = make_cooccurrence_matrix(conds[["201"]], print_matrix=T, heatmap_filename="201")
filt3e6l = make_cooccurrence_matrix(orders[["filt3E_6L"]])
fit_model <- function(model_name, orders, par_lower, par_upper) {
source(paste(model_dir,model_name,".R",sep=''))
par_init = (par_lower + par_upper) / 2
fits = list()
startt = Sys.time()
cat("Model\tOrder\tSSE\tParameters\n")
for(i in names(orders)) {
best <- psoptim(par_init, meanSSE, ord=orders[[i]]$train, human_perf=unlist(orders[[i]]$HumanItemAcc), lower=par_lower, upper=par_upper)
cat(model_name,'\t',i,'\t',best$value,'\t',best$par,'\n')
mod = model(best$par, ord=orders[[i]]$train)
fits[[names(orders)[i]]] = list(SSE=best$value, par=best$par, perf=mod$perf)
}
stopt = Sys.time()
print(stopt-startt)
return(fits)
}
meanSSE <- function(par, order, human_perf) {
mod = model(par, order)
# if there is an item grouping factor, can first aggregate item perf by the factor
#if(names(order$groups)!="all") { }
return(sum((mod$perf-human_perf)^2))
}
mod = run_model(orders[["freq369-3x3hiCD"]], "Bayesian_decay", c(.7,1,1), print_perf=T)
mod = run_model(conds[["211"]], "kachergis", c(1,3,.97), print_perf=T)
fit_model("kachergis", orders[1], c(.001,.1,.5), c(5,15,1))
fit_model("kachergis", orders[c(1,5,9)], c(.001,.1,.5), c(5,15,1))
fit_model("kachergis", conds, c(.001,.1,.5), c(5,15,1))
conds_with_data = c(paste(201:207),paste(215:225)) # 211-214 had 4afc test trials and so need special test code
conds_with_data = c(paste(201:207),"215","216","219",paste(221:225))
fit_model("kachergis", conds[conds_with_data], c(.001,.1,.5), c(5,15,1))
fit_model("strength", conds[conds_with_data], c(.001,.1), c(5,1))
fit_model("fazly", conds[conds_with_data], c(1e-10,5), c(.5,20000))
fit_model("Bayesian_decay", conds[conds_with_data], c(1e-5,1e-5,1e-5), c(10,10,10))
unc = fit_model("uncertainty", conds[conds_with_data], c(.001,.1,.5), c(5,15,1))
nov = fit_model("novelty", conds[conds_with_data], c(.001,.1,.5), c(5,15,1))
rwo = fit_model("rescorla-wagner", conds[conds_with_data], c(.0001,.1, .1), c(10,1,10))
rww = fit_model("rescorla-wagner_words_cues", conds[conds_with_data], c(.0001,.1, .1), c(10,1,10))
condnames = c("orig_3x3","freq369-3x3hiCD","freq369-3x3loCD","orig_4x4","cont_div6-12")
ka = fit_model("kachergis", orders[condnames], c(.001,.1,.5), c(5,15,1))
str = fit_model("strength", orders[condnames], c(.001,.1), c(5,1))
faz = fit_model("fazly", orders[condnames], c(1e-10,5,.1), c(.5,20000,1))
bay = fit_model("Bayesian_decay", orders[condnames], c(1e-5,1e-5,1e-5), c(10,10,10))
unc = fit_model("uncertainty", orders[condnames], c(.01,.5), c(15,1))
nov = fit_model("novelty", orders[condnames], c(.001,.01,.5), c(5,15,1))
rwo = fit_model("rescorla-wagner", orders[condnames], c(.0001,.01, .01), c(10,1,10))
rww = fit_model("rescorla-wagner_words_cues", orders[condnames], c(.0001,.01, .01), c(10,1,10))
#run_model(orders[["freq369-3x3hiCD"]], "Bayesian_decay", c(5.569798, 8.028788, 3.048987), print_perf=T)
faz_parms = c(.0001, 1000, .7)
f = run_model(orders[["freq369-3x3hiCD"]], "fazly", faz_parms, print_perf=T)
# according to Fazly et al. should be: .84 .94 .88 .75 .80 .89 .92 .95 .93 .98 .89 .92 .92 .98 .88 .88 .98 .85
faz_perf = c(.84, .94, .88, .75, .80, .89, .92, .95, .93, .98,.89,.92,.92,.98,.88,.88,.98,.85)
# freq3: .85 freq6: .93 freq9: .92
f = run_model(orders[["freq369-3x3hiCD"]], "fazly", faz_parms, print_perf=T)
cor(faz_perf, f$perf) # .97! really close...
# with no threshold, and with dummy word
fazf = fit_model("fazly", orders[condnames], c(1e-12,5,1e-12), c(.5,60000,1))
faza = fit_model("fazly", conds[conds_with_data], c(1e-12,5,1e-12), c(.5,60000,1))
# with threshold and with dummy word
fazf = fit_model("fazly", orders[condnames], c(1e-12,5,1e-12), c(.5,60000,1))
faza = fit_model("fazly", conds[conds_with_data], c(1e-12,5,1e-12), c(.5,60000,1))
f3 = run_model(orders[["freq369-3x3hiCD"]], "fazly", c(0.01228797, 169.3632, 0.6993908), print_perf=T)
mean(f3$perf[1:6]) # .43
mean(f3$perf[7:12]) # .55
mean(f3$perf[13:18]) # .65
# no threshold no dummy word
fazf = fit_model("fazly", orders[condnames], c(1e-12,5,1e-12), c(.5,60000,1))
faza = fit_model("fazly", conds[conds_with_data], c(1e-12,5,1e-12), c(.5,60000,1))
multinomial_likelihood_perfect <- function(par, ord) {
M = model(par, ord=ord)
pOgW = diag(M) / rowSums(M) # p(o|w)
lik = sum(log(pOgW))
return(-lik) # 18*log(1/18) = -52.02669 for AFC guessing
}
multinomial_likelihood <- function(cdat, M) {
M = M / rowSums(M) # p(o|w)
lik = 0
for(i in 1:dim(cdat)[1]) {
wordi = cdat[i,"Word"]
response_probs = M[wordi,unlist(cdat[i,c("Obj1","Obj2","Obj3","Obj4")])] # strengths of test objects
resp = cdat[wordi,]$Response #resp = cdat[which(cdat$Word==i),]$Response
lik = lik + log(M[wordi,resp] / sum(response_probs))
}
return(-lik) # minimize -loglik
}
binomial_likelihood <- function(cdat, M) {
est_prob = diag(M) / rowSums(M) # prob correct
lik = 0
for(i in 1:length(cdat)) { # dim(M)[1]
resp = cdat[i]
if(resp==i) {
lik = lik + log(est_prob[i])
} else {
lik = lik + log(1-est_prob[i])
}
}
return(lik) # 18*log(1/18) = -52.02669 for guessing
}
fit_subj <- function(par, ord, sdat) {
tot_lik = 0
M <- model(par, ord=ord)
# need: for each word (CorrectAns), what object they chose (Response)
mlik = binomial_likelihood(sdat, M)
return(-mlik)
}
fit_all <- function(par, ord, dat) {
tot_lik = 0
for(s in dim(dat)[1]) {
sdat <- unlist(dat[s,])
M <- model(par, ord=ord)
# need: for each word (CorrectAns), what object they chose (Response)
tot_lik = tot_lik + binomial_likelihood(sdat, M)
}
mlik = tot_lik #/ length(unique(dat$Subject))
return(-mlik)
}