-
Notifications
You must be signed in to change notification settings - Fork 0
/
actor_critic_lik_rt.asv
186 lines (157 loc) · 5.83 KB
/
actor_critic_lik_rt.asv
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
function [lik, latents] = actor_critic_lik_rt(x,DATA)
% Fit actor-critic agent model to data
%
% USAGE: simdata = actor_critic_lik(x,data)
agent.policy_update = DATA.m;
% fixed
agent.lrate_r = 0.01;
agent.lrate_e = 0.01;
agent.t0 = 150;
agent.beta0 = 1;
agent.sigma = 1;
agent.cost = 1;
agent.lrate_beta = 0;
if contains(agent.policy_update, 'fixed')
agent.C = [];
agent.V = [];
agent.beta0 = x(1);
agent.lrate_theta = x(2);
agent.lrate_V = x(3);
agent.lrate_p = x(4);
agent.b1 = x(5);
agent.b2 = x(6);
elseif contains(agent.policy_update, 'f_multibeta')
agent.C = [];
agent.V = [];
agent.beta0 = x(1);
agent.lrate_theta = x(7);
agent.lrate_V = x(8);
agent.lrate_p = x(9);
agent.b1 = x(10);
agent.b2 = x(11);
elseif contains(agent.policy_update, 'capacity')
agent.C = x(1);
agent.V = [];
agent.beta0 = x(2);
agent.lrate_theta = x(3);
agent.lrate_V = x(4);
agent.lrate_beta = x(5);
agent.lrate_p = x(6);
agent.compress = x(7);
agent.b1 = x(8);
agent.b2 = x(9);
elseif contains(agent.policy_update, 'value')
agent.C = [];
agent.V = x(1);
agent.beta0 = x(2);
agent.lrate_theta = x(3);
agent.lrate_V = x(4);
agent.lrate_beta = x(5);
agent.lrate_p = x(6);
agent.compress = x(7);
agent.b1 = x(8);
agent.b2 = x(9);
elseif contains(agent.policy_update, 'cv')
agent.C = x(1);
agent.V = x(2);
agent.beta0 = x(3);
agent.lrate_theta = x(4);
agent.lrate_V = x(5);
agent.lrate_beta = x(6);
agent.lrate_p = x(7);
agent.compress = x(8);
agent.b1 = x(9);
agent.b2 = x(10);
end % assign parameters
% so far best fit is t0=200 and fitting 10 parameters
lik = 0; latents.lik = zeros(1,length(DATA.exp)); k = 1;
exps = [1 2 3];
for eix = exps
data = DATA.exp(eix);
C = unique(data.cond);
lik_rt = zeros(length(C),sum(data.cond==1));
lik_choice = 0;
for c = 1:length(C)
if eix == 3 && c == 2 && isfield(agent,'compress') && contains(agent.policy_update,'c')
agent.C = agent.compress;
agent.lrate_p = agent.lrate_p + 0.1;
%if agent.lrate_p >1
% agent.lrate_p = 0.99;
%end
elseif eix == 3 && c == 2 && isfield(agent,'compress') && contains(agent.policy_update,'value')
agent.V = agent.compress;
agent.lrate_p = agent.lrate_p + 0.1;
%if agent.lrate_p >1
% agent.lrate_p = 0.99;
%end
end
ix = find(data.cond==C(c));
state = data.s(ix);
action = data.a(ix);
reward = data.r(ix);
rt = data.rt(ix);
R = data.Q(:,:,ix(1));
setsize = length(unique(state)); % number of distinct stimuli
nA = size(R,2); % number of distinct actions
V = zeros(setsize,1); % state values
theta = zeros(setsize,nA); % policy parameters
beta = agent.beta0;
if contains(agent.policy_update,'multibeta')
beta = x(k);
k = k+1;
end
p = ones(1,nA)/nA;
ecost = mutual_information(state(1:10),action(1:10),0.1);
rho = mean(R(:));
for t = 1:length(state)
s = state(t); a = action(t); r = reward(t);
if isnan(a) || a<1 || rt(t)<5
continue; % skip over non-responses
end
d = beta*theta(s,:) + log(p);
logpolicy = d - logsumexp(d);
policy = exp(logpolicy); % softmax policy
entropy = -nansum(policy.*log2(policy));
if logpolicy(a)<-10
logpolicy(a) = -10;
end
lik_choice = lik_choice + logpolicy(a);
cost = logpolicy(a) - log(p(a)); % policy complexity cost
rpe = beta*r - cost - V(s); % reward prediction error
lik_rt(c,t) = (log(rt(t)) - log(agent.t0 + agent.b1*abs(cost) + agent.b2*entropy))^2;
rho = rho + agent.lrate_r*(r-rho); % avg reward update
ecost = ecost + agent.lrate_e*(cost-ecost); % policy cost update
chosen = a; idxs = 1:nA; unchosen = idxs(idxs~=chosen);
g(:,chosen) = beta*(1-policy(chosen)); % policy gradient for chosen actions
g(:,unchosen) = beta*(-policy(unchosen)); % policy gradient for unchosen actions
theta(s,:) = theta(s,:) + agent.lrate_theta*rpe*g; % policy parameter update
V(s) = V(s) + agent.lrate_V*rpe;
if agent.lrate_beta > 0
switch agent.policy_update
case 'capacity'
beta = beta + agent.lrate_beta*(agent.C-ecost);
case 'value'
beta = beta + agent.lrate_beta*(agent.V-rho);
case 'cv'
beta = beta + agent.lrate_beta*((agent.C-ecost)/(agent.V-rho)-beta);
case 'pg'
beta = beta + agent.lrate_beta*rpe*(theta(s,a)-(theta(s,:)*policy'));
case 'capacity_p'
beta = beta + agent.lrate_beta*(agent.C-ecost);
case 'value'
beta = beta + agent.lrate_beta*(agent.V-rho);
end
beta = max(min(beta,30),0.1);
end
if agent.lrate_p > 0
p = p + agent.lrate_p*(policy - p); p = p./sum(p); % marginal update
end
end
n(c) = length(state);
end % each condition
lik_rt(lik_rt<-10) = -10;
lik_rt(lik_rt>1e5) = 1e5;
lik = lik + lik_choice + sum(n)*log(1/(sqrt(2*pi)*agent.sigma))-(1/(2*agent.sigma^2))*sum(lik_rt(:));
latents.lik(eix) = latents.lik(eix) + lik_choice + sum(n)*log(1/(sqrt(2*pi)*agent.sigma))-(1/(2*agent.sigma^2))*sum(lik_rt(:));
end % each experiment
end % function