-
Notifications
You must be signed in to change notification settings - Fork 4
/
knowledgeGradientHumanPolicy.py
139 lines (121 loc) · 6.42 KB
/
knowledgeGradientHumanPolicy.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
from __future__ import print_function, division
import numpy as np
import tensorflow as tf
import gym
from gym.utils import seeding
from gym import spaces
from sandbox.rocky.tf.spaces.discrete import Discrete #supercedes the gym spaces
from assistive_bandits.envs.utils import softmax
from assistive_bandits.envs.humanPolicy.humanPolicy import HumanPolicy
class KGBetaBernoulliBanditPolicy(HumanPolicy):
"""
Implements a knowledge gradient-based policy for a Beta-Bernoulli Bandit.
Decision rule is:
argmax_k (theta^t_k + (T - t -1) v^KG,t_k)
where v^KG is the knowledge gradient.
"""
def __init__(self, env):
assert (isinstance(env.observation_space, Discrete) \
or isinstance(env.observation_space, spaces.Discrete))
assert (isinstance(env.action_space, Discrete) \
or isinstance(env.action_space, spaces.Discrete))
self.T = env.horizon
super(KGBetaBernoulliBanditPolicy, self).__init__(env)
def reset(self):
self.t = 0
self.successes = np.ones(self.env.nA)
self.failures = np.ones(self.env.nA)
self.means = self.successes/(self.successes + self.failures)
self.knowledge_gradient = np.zeros(self.env.nA)
for i in range(self.env.nA):
new_means_1 = self.means.copy()
new_means_2 = self.means.copy()
new_means_1[i] = (self.successes[i] + 1)/(self.successes[i] + self.failures[i] + 1)
new_means_2[i] = (self.successes[i])/(self.successes[i] + self.failures[i] + 1)
expected_new_max_mean = self.means[i] * np.max(new_means_1) + (1-self.means[i])*np.max(new_means_2)
self.knowledge_gradient[i] = expected_new_max_mean
self.knowledge_gradient -= np.max(self.means)
def get_action(self, obs):
self.arm_vals = self.means + (self.T - self.t + 1) * self.knowledge_gradient
return self.np_random.choice(np.where(np.isclose(self.arm_vals, np.max(self.arm_vals)))[0])
def learn(self, old_obs, act, rew, new_obs, done):
self.t += 1
if rew > 0.5:
self.successes[act] += 1
else:
self.failures[act] += 1
self.means[act] = self.successes[act]/(self.successes[act] + self.failures[act])
#update knowledge gradient
for i in range(self.env.nA):
new_means_1 = self.means.copy()
new_means_2 = self.means.copy()
new_means_1[i] = (self.successes[i] + 1)/(self.successes[i] + self.failures[i] + 1)
new_means_2[i] = (self.successes[i])/(self.successes[i] + self.failures[i] + 1)
expected_new_max_mean = self.means[i] * np.max(new_means_1) + (1-self.means[i])*np.max(new_means_2)
self.knowledge_gradient[i] = expected_new_max_mean
self.knowledge_gradient -= np.max(self.means)
#Functions for MCMC
def get_initial_state(self, **kwargs):
t = 0
successes = np.ones(self.env.nA)
failures = np.ones(self.env.nA)
means = successes/(successes+failures)
knowledge_gradient = np.zeros(self.env.nA)
for i in range(self.env.nA):
new_means_1 = self.means.copy()
new_means_2 = self.means.copy()
new_means_1[i] = (self.successes[i] + 1)/(self.successes[i] + self.failures[i] + 1)
new_means_2[i] = (self.successes[i])/(self.successes[i] + self.failures[i] + 1)
expected_new_max_mean = self.means[i] * np.max(new_means_1) + (1-self.means[i])*np.max(new_means_2)
knowledge_gradient[i] = expected_new_max_mean
knowledge_gradient -= np.max(means)
return {'t': t, 'means': means, 'successes': successes, 'failures': failures, 'kg': knowledge_gradient}
def get_action_from_state(self, state, obs):
arm_vals = state['means'] + (self.T - self.t + 1) * state['kg']
act = self.np_random.choice(np.where(np.isclose(arm_vals, arm_vals.max()))[0])
return act
def update_state(self, state, old_obs, act, rew, new_obs):
state['t'] += 1
if rew > 0.5:
state['successes'][act] += 1
else:
state['failures'][act] += 1
state['means'][act] = state['successes'][act]/(state['successes'][act] + state['failures'][act])
#update knowledge gradient
for i in range(self.env.nA):
new_means_1 = state['means'].copy()
new_means_2 = state['means'].copy()
new_means_1[i] = (state['successes'][i] + 1)/(state['successes'][i] + state['failures'][i] + 1)
new_means_2[i] = (state['successes'][i])/(state['successes'][i] + state['failures'][i] + 1)
expected_new_max_mean = state['means'][i] * np.max(new_means_1) + (1-state['means'][i])*np.max(new_means_2)
state['kg'][i] = expected_new_max_mean
state['kg'] -= np.max(state['means'])
return state
def likelihood(self, state, obs, act):
""" Computes the likelihood of taking the action given the state. """
arm_vals = state['means'] + (self.T - self.t + 1) * state['kg']
greedy_arms = np.where(np.isclose(arm_vals, arm_vals.max()))[0]
return 1/len(greedy_arms) if act in greedy_arms else 1e-10
def log_likelihood(self, state, obs, act):
return np.log(self.likelihood(state, obs, act))
#RTDP
def act_probs_from_counts(self, counts, *args, **kwargs):
t = sum(counts) - 2*self.env.nA
successes = np.array([counts[2*i] for i in range(self.env.nA)])
failures = np.array([counts[2*i+1] for i in range(self.env.nA)])
means = successes/(successes + failures)
#calculate knowledge gradient
knowledge_gradient = np.zeros(self.env.nA)
for i in range(self.env.nA):
new_means_1 = means.copy()
new_means_2 = means.copy()
new_means_1[i] = (successes[i] + 1)/(successes[i] + failures[i] + 1)
new_means_2[i] = (successes[i])/(successes[i] + failures[i] + 1)
expected_new_max_mean = means[i] * np.max(new_means_1) + (1-means[i])*np.max(new_means_2)
knowledge_gradient[i] = expected_new_max_mean
knowledge_gradient -= np.max(means)
arm_vals = means + (self.T - t + 1) * knowledge_gradient
act_probs = np.zeros(self.env.nA, dtype=np.float32)
greedy_arms = np.where(np.isclose(arm_vals,arm_vals.max()))[0]
act_probs[greedy_arms] += 1/len(greedy_arms)
return act_probs