def __init__(self):
super().__init__(nactions=4,
noutcomes=3,
nstates=4,
min_actions_per_context=None,
alpha=0.1,
alpha_start=1.,
shift_flip='shift',
reward_lb=-1,
reward_ub=1,
reward_drift='on',
drift_mu=np.zeros(3),
drift_sd=1.)
data = generate_behavioural_data(MyBanditTask, RWSoftmaxAgent, 20, 200)
def log_prob(w, D):
agent = RWSoftmaxAgent(task=MyBanditTask(),
learning_rate=w[0],
inverse_softmax_temp=w[1])
L = 0
for t in range(D.shape[0]):
x = D[t, :7]
u = D[t, 7:11]
r = D[t, 11]
x_ = D[t, 12:]
L += u @ agent.log_prob(x)
agent.learning(x, u, r, x_, None)
return L
def test_cooccurrence_matrix():
data1 = generate_behavioural_data(TwoArmedBandit, RWSoftmaxAgent, 10, 20)
data1.make_behavioural_ngrams(2)
data1.make_cooccurrence_matrix(2)
def test_bdf():
data1 = generate_behavioural_data(TwoArmedBandit, RWSoftmaxAgent, 10, 20)
data2 = generate_behavioural_data(TwoArmedBandit, RWSoftmaxAgent, 10, 20)
data = merge_behavioural_data([data1, data2])
import numpy as np
import matplotlib.pyplot as plt
from fitr import generate_behavioural_data
from fitr.environments import TwoArmedBandit
from fitr.agents import RWSoftmaxAgent
from fitr.inference import mlepar
from fitr.utils import sigmoid
from fitr.utils import stable_exp
from fitr.criticism.plotting import actual_estimate
import fitr.gradients as grad
N = 50 # number of subjects
T = 200 # number of trials
# Generate synthetic data
data = generate_behavioural_data(TwoArmedBandit, RWSoftmaxAgent, N, T)
# Create log-likelihood function
def log_prob(w, D):
lr = sigmoid(w[0], a_min=-6, a_max=6)
ist = stable_exp(w[1], a_min=-10, a_max=10)
agent = RWSoftmaxAgent(TwoArmedBandit(), lr, ist)
L = 0
for t in range(D.shape[0]):
x = D[t, :3]
u = D[t, 3:5]
r = D[t, 5]
x_ = D[t, 6:]
agent.log_prob(x, u)
agent.learning(x, u, r, x_, None)