-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodels.py
340 lines (291 loc) · 12.4 KB
/
models.py
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
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
from collections import OrderedDict as odict
from lmfit import Minimizer, Parameters
import numpy as np
import pandas as pd
# Functions to define Shift Task model (Wilson & Niv, 2012)
class BasicRLModel(object):
""" Basic RL Model for RL task """
def __init__(self, data, verbose=False):
self.data = data
self.init_val = .5
self.stims = np.unique(data.stim_indices.apply(lambda x: x[0]))
self.vals = odict({i: self.init_val for i in self.stims})
# set up class vars
self.beta=1
self.eps=0
self.lr = .01
self.verbose=verbose
def _calc_choice_probs(self, trial):
stims = trial['stim_indices']
stim_values = [self._get_value(stim) for stim in stims]
# compute softmax decision probs
f = lambda x: np.e**(self.beta*x)
softmax_values = [f(v) for v in stim_values]
normalized = [v/np.sum(softmax_values) for v in softmax_values]
return normalized
def _get_choice_prob(self, trial):
probs = self._calc_choice_probs(trial)
# get prob of choice
choice_prob = probs[int(trial['response'])]
# incorporate eps
choice_prob = (1-self.eps)*choice_prob + (self.eps)*(1/3)
return choice_prob
def _get_value(self, stim):
return self.vals[stim]
def _reset_vals(self):
for k in self.vals:
self.vals[k] = self.init_val
def _update(self, trial):
selected_stim = trial['selected_stim']
reward = trial['rewarded']
value = self._get_value(int(selected_stim))
delta = self.lr*(reward-value)
self.vals[selected_stim] += delta
def _wipe_data(self, data):
cols = ['correct', 'response','rewarded', 'rt',
'selected_stim', 'selected_value']
for col in cols:
if col in data.columns:
data.loc[:, col] = np.nan
def get_log_likelihood(self, start=None, stop=None):
""" Returns summed negative log likelihood """
probs, track_vals = self.run_data()
if start:
probs = probs[start:]
if stop:
probs = probs[:stop]
neg_log_likelihood = -np.sum(np.log(probs))
return neg_log_likelihood
def get_params(self):
return {'beta': self.beta,
'lr': self.lr,
'eps': self.eps}
def update_params(self, params):
self.__dict__.update(params)
def run_data(self):
self._reset_vals()
probs = []
track_vals = []
for i, trial in self.data.iterrows():
probs.append(self._get_choice_prob(trial))
self._update(trial)
track_vals.append(self.vals.copy())
return probs, track_vals
def simulate(self, prob_match=True, data=None):
if data is None:
data = self.data
self._reset_vals()
sim_data = []
for i, trial in data.iterrows():
trial = trial.to_dict()
# make decision
probs = self._calc_choice_probs(trial)
if prob_match:
choice = np.random.choice([0,1], p=probs)
else:
choice = np.argmax(probs)
trial['response_prob'] = probs[choice]
trial['response'] = choice
trial['rewarded'] = trial['rewards'][choice]
trial['selected_stim'] = trial['stim_indices'][choice]
trial['selected_value'] = trial['values'][choice]
if np.isnan(trial['correct_choice']) == False:
trial['correct'] = choice==trial['correct_choice']
else:
trial['correct'] = np.nan
sim_data.append(trial)
# update
if trial['display_reward']:
self._update(trial)
# change type of columns
sim_data = pd.DataFrame(sim_data)
sim_data.selected_stim = sim_data.selected_stim.astype(int)
sim_data.correct = sim_data.correct.astype(float)
return sim_data
def optimize(self, start=None, stop=None):
def loss(pars):
#unpack params
parvals = pars.valuesdict()
self.beta = parvals['beta']
self.lr = parvals['lr']
self.eps = parvals['eps']
return self.get_log_likelihood(start, stop)
def track_loss(params, iter, resid):
if iter%100==0:
print(iter, resid, params)
params = Parameters()
params.add('beta', value=1, min=.01, max=100)
params.add('eps', value=0, min=0, max=1)
params.add('lr', value=.1, min=.000001, max=1)
if self.verbose==False:
fitter = Minimizer(loss, params)
else:
fitter = Minimizer(loss, params, iter_cb=track_loss)
fitter.scalar_minimize(method='Nelder-Mead', options={'xatol': 1e-3,
'maxiter': 200})
def optimize_acc(self, data=None):
def loss(pars):
#unpack params
parvals = pars.valuesdict()
self.beta = parvals['beta']
self.lr = parvals['lr']
self.eps = parvals['eps']
sim_data = self.simulate(data=data, prob_match=True)
# get probability of correct response
prob_correct = abs((1-sim_data.correct)-sim_data.response_prob)
return -np.sum(np.log(prob_correct))
def track_loss(params, iter, resid):
if iter%100==0:
print(iter, resid, params)
params = Parameters()
params.add('beta', value=1, min=.01, max=100)
params.add('eps', value=0, min=0, max=1)
params.add('lr', value=.1, min=.000001, max=1)
if self.verbose==False:
fitter = Minimizer(loss, params)
else:
fitter = Minimizer(loss, params, iter_cb=track_loss)
fitter.scalar_minimize(method='Nelder-Mead', options={'xatol': 1e-3,
'maxiter': 200})
def SR_from_transition(transitions, gamma):
""" Computes successor representation from one-step transition matrix """
I = np.identity(transitions.shape[1])
M=np.linalg.inv(I-gamma*transitions)
return M
class SR_RLModel(BasicRLModel):
""" SR-RL Model
Russek, E. M., et al (2017). Predictive representations can link model-based reinforcement learning to model-free mechanisms.
"""
def __init__(self, RLdata, StructureData, verbose=False):
super(SR_RLModel, self).__init__(RLdata, verbose)
self.structure = StructureData
self.SR_lr = .05
self.gamma = .99
self.M = self.SR_TD()
def _get_value(self, stim):
return self.M[stim,:].dot(list(self.vals.values()))/np.sum(self.M[stim,:])
def get_M(self):
return self.M
def get_params(self):
params = super(SR_RLModel, self).get_params()
params['SR_lr'] = self.SR_lr
params['gamma'] = self.gamma
return params
def optimize(self, start=None, stop=None):
def loss(pars):
#unpack params
parvals = pars.valuesdict()
# base RL params
self.beta = parvals['beta']
self.eps = parvals['eps']
self.lr = parvals['lr']
# SR params
self.gamma = parvals['gamma']
self.SR_lr = parvals['SR_lr']
self.M = self.SR_TD()
return self.get_log_likelihood(start, stop)
def track_loss(params, iter, resid):
if iter%100==0:
print(iter, resid)
params = Parameters()
params.add('beta', value=1, min=.01, max=100)
params.add('eps', value=0, min=0, max=1)
params.add('lr', value=.1, min=.000001, max=1)
params.add('gamma', value=1, min=0, max=1)
params.add('SR_lr', value=0, min=0, max=1)
if self.verbose==False:
fitter = Minimizer(loss, params)
else:
fitter = Minimizer(loss, params, iter_cb=track_loss)
fitter.scalar_minimize(method='Nelder-Mead', options={'xatol': 1e-3,
'maxiter': 200})
def optimize_acc(self, data=None):
def loss(pars):
#unpack params
parvals = pars.valuesdict()
# base RL params
self.beta = parvals['beta']
self.eps = parvals['eps']
self.lr = parvals['lr']
# SR params
self.gamma = parvals['gamma']
self.SR_lr = parvals['SR_lr']
self.M = self.SR_TD()
sim_data = self.simulate(data=data, prob_match=True)
prob_correct = abs((1-sim_data.correct)-sim_data.response_prob)
return -np.sum(np.log(prob_correct))
def track_loss(params, iter, resid):
if iter%100==0:
print(iter, resid)
params = Parameters()
params.add('beta', value=1, min=.01, max=100)
params.add('eps', value=0, min=0, max=1)
params.add('lr', value=.1, min=.000001, max=1)
params.add('gamma', value=1, min=0, max=1)
params.add('SR_lr', value=.05, min=0, max=1)
if self.verbose==False:
fitter = Minimizer(loss, params)
else:
fitter = Minimizer(loss, params, iter_cb=track_loss)
fitter.scalar_minimize(method='Nelder-Mead', options={'xatol': 1e-3,
'maxiter': 200})
def simulate(self, prob_match=True, data=None):
self.M = self.SR_TD()
sim_data = super(SR_RLModel, self).simulate(prob_match, data)
return sim_data
def SR_TD(self):
state_df = pd.DataFrame({'s': self.structure.stim_index,
'sprime': self.structure.stim_index.shift(-1)}).iloc[:-1].astype(int)
states = state_df.s.unique()
M = np.identity(len(states)) # define future-state occupancy matrix
def _update(s, sprime):
base = np.zeros(M.shape[1]); base[s]=1
delta = base+self.gamma*M[sprime,:]-M[s,:]
M[s,:] += self.SR_lr*delta
for i, trial in state_df.iterrows():
_update(trial['s'], trial['sprime'])
return M
class GraphRLModel(BasicRLModel):
""" Extends basic RL Model with value propagation one edge in graph """
def __init__(self, data, graph, verbose=False):
super(GraphRLModel, self).__init__(data, verbose)
self.graph = graph
self.graph_decay = 0 # if 0, no propagation
def _update(self, trial):
selected_stim = trial['selected_stim']
reward = trial['rewarded']
value = self.vals[selected_stim]
delta = self.lr*(reward-value)
self.vals[selected_stim] += delta
# value propagation
for stim in self.graph[selected_stim]:
value = self.vals[stim]
delta = self.lr*(reward-value)*self.graph_decay
self.vals[stim] += delta
def get_params(self):
params = super(GraphRLModel, self).get_params()
params['graphydecay'] = self.graph_decay
return params
def optimize(self, start=None, stop=None):
def loss(pars):
#unpack params
parvals = pars.valuesdict()
self.beta = parvals['beta']
self.eps = parvals['eps']
self.graph_decay = parvals['graph_decay']
self.lr = parvals['lr']
return self.get_log_likelihood(start, stop)
def track_loss(params, iter, resid):
if iter%100==0:
print(iter, resid)
params = Parameters()
params.add('beta', value=1, min=.01, max=100)
params.add('eps', value=0, min=0, max=1)
params.add('graph_decay', value=0, min=0, max=1)
params.add('lr', value=.1, min=.000001, max=1)
if self.verbose==False:
fitter = Minimizer(loss, params)
else:
fitter = Minimizer(loss, params, iter_cb=track_loss)
fitter.scalar_minimize(method='Nelder-Mead', options={'xatol': 1e-3,
'maxiter': 200})