PF_MAB/bandits_improved.py at main · ShenGroup/PF_MAB · GitHub

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
# -*- coding: utf-8 -*-

import numpy as np
from client_improved import imp_client
from server_improved import imp_server

class IMP_PFEDUCB(object):
    def __init__(self,
                 fun_fp,
                 T,
                 means, # M*K
                 alpha,
                 reward='Gaussian'):
        self.M = means.shape[0]
        self.K = means.shape[1]
        self.local_means = means
        self.reward_type = reward
        self.alpha = alpha
        self.T = T
        self.C = 1
        self.comm = 0

        self.global_means = np.sum(self.local_means, axis=0)/self.M
        self.clients = [
            imp_client(index = i, thorizon = self.T, narms=self.K, nclients = self.M, palpha = self.alpha, fp=fun_fp) for i in range(self.M)
        ]
        self.server = imp_server(narms=self.K, nclients = self.M)


    def simulate_single_step_rewards(self):
        if self.reward_type == 'Bernoulli':
            return np.random.binomial(1, self.local_means)
        return np.random.normal(self.local_means, 1)

    def simulate_single_step(self, plays):

        local_rews = self.simulate_single_step_rewards()
        global_rews = np.mean(local_rews,axis=0)

        local_rewards = np.array([local_rews[i,plays[i]] for i in range(self.M)])
        global_rewards = np.array([global_rews[plays[i]] for i in range(self.M)])
        mixed_rewards = self.alpha*local_rewards+(1-self.alpha)*global_rewards
        #rewards = np.array([self.alpha*local_rews[i,plays[i]]+(1-self.alpha)*global_rews[plays[i]] for i in range(self.M)])

        return local_rewards, global_rewards, mixed_rewards

    def simulate(self):
        """
        Return the vector of regret for each time step until horizon
        """

        local_rewards = []
        global_rewards = []
        mixed_rewards = []

        play_history = []


        for t in range(self.T):
            plays = np.zeros(self.M)

            plays = [(int)(client.play()) for client in self.clients]
            local_rews, global_rews, mixed_rews = self.simulate_single_step(plays)
            #obs, rews = self.simulate_single_step(plays)  # observations of all players

            local_rewards.append(np.sum(local_rews))
            global_rewards.append(np.sum(global_rews))
            mixed_rewards.append(np.sum(mixed_rews))
            play_history.append(plays)

            #print(plays)

            for i in range(self.M):
                self.clients[i].reward_update(plays[i], local_rews[i])  # update strategies of all player

            for i in range(self.M):
                f_local_stat, local_stat = self.clients[i].local_mean_update()
                if f_local_stat is True:
                    self.server.local_mean_update(i,local_stat)
                f_local_stat = False

            f_global_stat, global_stat = self.server.global_mean_update()

            if f_global_stat is True:
                for i in range(self.M):
                    self.clients[i].global_mean_update(global_stat)
                    local_set = self.clients[i].local_set_update()
                    local_delta = self.clients[i].local_delta_update()
                    self.server.local_set_update(i,local_set)
                    self.server.local_delta_update(i,local_delta)

                global_set = self.server.global_set_update()
                global_delta = self.server.global_delta_update()
                #print(t," global-set:",global_set)
                for i in range(self.M):
                    self.clients[i].global_set_update(global_set)
                    self.clients[i].global_delta_update(global_delta)
                    local_rewards[-1] -= 2*self.C
                    global_rewards[-1] -= 2*self.C
                    mixed_rewards[-1] -= 2*self.C
                    self.comm += 2*self.C
                    #print(t,"comm")
                f_global_stat = False

        top_mixed_means = np.zeros(self.M)

        top_arms = np.zeros(self.M, dtype=int)
        sub_gap = np.zeros(self.M)
        for i in range(self.M):
            top_arms[i] = np.argmax(self.alpha*self.local_means[i]+(1-self.alpha)*self.global_means)
            top_mixed_means[i] = self.alpha*self.local_means[i][top_arms[i]]+(1-self.alpha)*self.global_means[top_arms[i]]
            sub_gap[i] = np.sort(self.alpha*self.local_means[i]+(1-self.alpha)*self.global_means)[-2]-top_mixed_means[i]
            #print("player", i, " top_arm:",top_arms[i], " top_mean:", top_mixed_means[i], " gap:", sub_gap[i])

        best_case_reward = np.sum(top_mixed_means) * np.arange(1, self.T + 1)

        cumulated_local_reward = np.cumsum(local_rewards)
        cumulated_global_reward = np.cumsum(global_rewards)
        cumulated_mixed_reward = np.cumsum(mixed_rewards)

        regret = best_case_reward - cumulated_mixed_reward
#        print(cumulated_mixed_reward)
#        print(best_case_reward)
#        print('regret:',regret[-1], "comm: ", self.comm)
        self.regret = (regret, best_case_reward, cumulated_mixed_reward)
        self.top_mixed_means = top_mixed_means
        return regret, cumulated_local_reward, cumulated_global_reward, cumulated_mixed_reward

    def get_clients(self):
        return self.clients