forked from cstarkweather/Belief-State-TD-Model
-
Notifications
You must be signed in to change notification settings - Fork 0
/
WithOmission.m
260 lines (220 loc) · 7.94 KB
/
WithOmission.m
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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
% Dopamine RPEs Reflect Hidden State Inference Across Time
% Task 2 Model (90% Rewarded)
% POMDP in the spirit of Daw et al. (2006), Rao (2010)
% Authors: Clara Kwon Starkweather, Dr. Samuel J. Gershman
%initialize
nTrials = 1000; %number of trials
x=[]; %series of observations during trials - will be filled in later
% Gaussian probability distribution and cumulative probability distribution
ISIpdf=normpdf([1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8],2,0.5)/sum(normpdf([1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8],2,0.5));
ISIcdf=[0.0477 0.1312 0.2558 0.4142 0.5858 0.7442 0.8688 0.9523 1.0000];
% Uncomment 2 lines directly below for flat probability distribution -
% as seen in Supplementary Figure 9
% ISIpdf=[0.111 0.111 0.111 0.111 0.111 0.111 0.111 0.111 0.111];
% ISIcdf=[0.111 0.222 0.333 0.444 0.555 0.666 0.777 0.888 1.0];
% Create distribution of ISI's - 10% unassigned are omission trials
% Possible ISIs range from 5-13
for i=1:nTrials-nTrials/10
ISIdistributionMatrix(i)=sum(ISIcdf<rand)+5;
end
ISIdistributionMatrix=[ISIdistributionMatrix NaN(1,nTrials/10)]; %omission trials are marked as 'NaN'
ISIdistributionMatrix=ISIdistributionMatrix(randperm(length(ISIdistributionMatrix))); %Randomize when the omission trials occur
% Calculate hazard rate of receiving reward after substates 5-14 (ISIhazard);
% Used later to create transition matrix
ISIhazard(1)=ISIpdf(1);
for i=2:length(ISIpdf)
ISIhazard(i)=ISIpdf(i)/(1-ISIcdf(i-1));
end
% Set hazard rate of transitioning OUT of the ITI
% 1/65 is based on task parameters
ITIhazard=(1/65);
% Generate sequence of observations that corresponds to trials
% Observations:
% Null-->1
% Odor ON-->2
% Reward-->3
for i=1:nTrials
if ~isnan(ISIdistributionMatrix(i)) %reward delivery trials
ISI = ones(1,ISIdistributionMatrix(i));
ITI = ones(1,geornd(ITIhazard));
trial=[2;ISI';3;ITI'];
x=[x;trial];
else %omission trials
ITI = ones(1,geornd(ITIhazard));
trial=[2;ITI'];
x=[x;trial];
end
ITIdistributionMatrix(i)=length(ITI);
end
% states:
% ISI = 1-14
% ITI = 15
%Fill out the observation matrix O
% O(x,y,:) = [a b c]
% a is the probability that observation 1 (null) was observed given that a
% transition from sub-state x-->y just occurred
% b is the probability that observation 2 (odor ON) was observed given that
% a transition from sub-state x-->y just occurred
% c is the probability that observation 2 (reward) was observed given that
% a transition from sub-state x-->y just occurred
O=zeros(15,15,3);
%ISI
O(1,2,:) = [1 0 0];
O(2,3,:) = [1 0 0];
O(3,4,:) = [1 0 0];
O(4,5,:) = [1 0 0];
O(5,6,:) = [1 0 0];
O(6,7,:) = [1 0 0];
O(7,8,:) = [1 0 0];
O(8,9,:) = [1 0 0];
O(9,10,:) = [1 0 0];
O(10,11,:) = [1 0 0];
O(11,12,:) = [1 0 0];
O(12,13,:) = [1 0 0];
O(13,14,:) = [1 0 0];
%obtaining reward
O(14,15,:) = [0 0 1];
O(13,15,:) = [0 0 1];
O(12,15,:) = [0 0 1];
O(11,15,:) = [0 0 1];
O(10,15,:) = [0 0 1];
O(9,15,:) = [0 0 1];
O(8,15,:) = [0 0 1];
O(7,15,:) = [0 0 1];
O(6,15,:) = [0 0 1];
%stimulus onset
O(15,1,:) = [0 1 0]; %rewarded trial
O(15,15,2) = ITIhazard*0.1; %omission trial
%ITI
O(15,15,1) = 1-(ITIhazard*0.1);
%Fill out the transition matrix T
%T(x,y) is the probability of transitioning from sub-state x-->y
T=zeros(15,15);
%odor ON from substates 1-6
%no probability of transitioning out of ISI while odor ON
T(1,2)=1;
T(2,3)=1;
T(3,4)=1;
T(4,5)=1;
T(5,6)=1;
%T(ISIsubstate_i+6-->ISIsubstate_i+7) = ISIhazard(i)
%these substates span the variable ISI interval
%if reward is received, then transition into the ITI
for i=5:length(ISIhazard)+4
T(1+i,2+i)=1-ISIhazard(i-4);
T(1+i,15)=ISIhazard(i-4);
end
T(14,15)=1;
% ITI length is drawn from exponential distribution in task
% this is captured with single ITI substate with high self-transition
% probability
T(15,15)=1-(ITIhazard*0.9);
T(15,1)=ITIhazard*0.9;
%% Visualize the transition and observation matrices
%Code for Supplementary Figure 7
subplot(2,1,1)
imagesc(T)
title('Transition Matrix')
xlabel('Next Substate')
ylabel('Current Substate')
subplot(2,1,2)
imagesc(O)
title('Observation Matrix')
xlabel('Next Substate')
ylabel('Current Substate')
%% Run TD learning
results = TD(x,O,T);
%% plot RPE as a function of ISI; plot every trial
RewardIndices=find(x==3);
RewardIndices=RewardIndices(length(RewardIndices)*0.4:end); % only look at trials after 2000 trials
ISIdistributionMatrix_rewardedtrials=ISIdistributionMatrix(~isnan(ISIdistributionMatrix));
ISIsforplot=ISIdistributionMatrix_rewardedtrials(length(ISIdistributionMatrix_rewardedtrials)*0.4:end); % only look at trials after 2000 trials
plot(ISIsforplot,results.rpe(RewardIndices),'k*')
xlabel('ISI')
ylabel('TD error')
%% plot average RPE for each ISI
% Code for Supplementary Figure 5
RewardRPE=results.rpe(RewardIndices);
% Average RPEs (and standard error) for each ISI length
for i=1:9
averageRPE(i)=sum(RewardRPE(find(ISIsforplot==i+4)))/length(find(ISIsforplot==i+4));
errorRPE(i)=std(RewardRPE(find(ISIsforplot==i+4)))/sqrt(length(find(ISIsforplot==i+4)));
end
% Plotting average RPE and standard error for each ISI
for i=1:9
errorbar(i, averageRPE(i), errorRPE(i),'k')
hold on
plot(i, averageRPE(i), '.','Color',[1-i*.1 i*.1 1],'markersize',25)
hold on
end
xlabel('time of reward delivery','fontSize',20)
ylabel('Average TD error','fontSize',20)
%% Value, valueprime, and RPE
%Code for value signals and RPEs shown in Figure 6
clear cueonsets;
cueonsets=find(x==2);
whichISI=13; %how long is the ISI for the trial type that you want to plot the value signal for? range: 5-13
cueonsets=cueonsets(ISIdistributionMatrix==whichISI);
cueonsets=cueonsets(round(length(cueonsets)*0.4):end-12); % only look at trials after 2000 trials
value=zeros(1,20);
valueprime=zeros(1,20);
rpe=zeros(1,20);
for i=1:length(cueonsets)
for j=1:20
value(j)=results.value(cueonsets(i)+j-2)+value(j);
valueprime(j)=results.value(cueonsets(i)+j-1)+valueprime(j);
rpe(j)=results.rpe(cueonsets(i)+j-2)+rpe(j);
end
end
% plot value, value(t+1) and rpe
subplot(3,1,1)
plot(value/length(cueonsets),'k')
hold on
plot(valueprime/length(cueonsets),'Color',[0.5 0.5 0.5])
title('Value [black] and Value(t+1) [grey]')
subplot(3,1,2)
plot((valueprime-value)/length(cueonsets),'k')
title('Value(t+1)-Value(t)')
subplot(3,1,3)
plot(rpe/length(cueonsets),'Color',[1-(whichISI-4)/9 (whichISI-4)/9 1])
title('TD error')
%% Plot substate weights
%Code for weights shown in Figure 6
for i=1:15
weight(i)=sum(results.w(round(length(results.w)*0.4):end,i));
end
bar(weight/(round(length(results.w)*0.6)))
ylim([0 2])
ylabel('Weight')
xlabel('substate')
%% Generate matrix of RPEs for plotting
%Code for RPEs shown in Figure 5
RPE=zeros(10,30);
allOdorIndices=find(x==2);
odorIndicesforplot=allOdorIndices(length(allOdorIndices)*0.4:end-12);
ISIsforplot=ISIdistributionMatrix(length(allOdorIndices)*0.4:end-12);
for i=1:length(ISIsforplot)
if ~isnan(ISIsforplot(i))
RPE(ISIsforplot(i)-4,:)=(results.rpe((odorIndicesforplot(i)-5:odorIndicesforplot(i)+24)))'+RPE(ISIsforplot(i)-4,:);
else
end
end
% divide by the number of trials to compute each ISI's average RPE
for i=1:size(RPE,1)
RPE(i,:)=RPE(i,:)/sum(ISIsforplot==i+4);
end
plot((RPE(1:9,:))')
ylabel('TD Error')
xlabel('time')
%% Plot omission trials
% Code for Supplementary Figure 1
odorIndices=find(x==2);
Omissiontrials=odorIndices(isnan(ISIdistributionMatrix));
Omissiontrials=Omissiontrials(round(0.4*length(Omissiontrials)):end);
omissionRPEs=zeros(1,30);
for i=1:length(Omissiontrials)
omissionRPEs(1:30)=omissionRPEs(1:30)+(results.rpe(Omissiontrials(i)-4:Omissiontrials(i)+25))';
end
omissionRPEs=omissionRPEs/length(Omissiontrials);
plot(-0.8:0.2:5,omissionRPEs)
ylim([-0.8 0.8])