RTDP/rtdp.py at master · instance01/RTDP · 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
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
from collections import defaultdict

import numpy as np
from gym.envs.toy_text.frozen_lake import FrozenLakeEnv


MAP_20x20 = [
    'SFFFFFFFHFFFFFFHFHFF',
    'FFFFFFFFFFFHFFFFFHFF',
    'FFHFFHFHFFFFFFHFFFFH',
    'FFHFFHFFFFFFFFHFFHFF',
    'FFFHFFFFFFFFFFFFFFHF',
    'FFFFHFFFFFHFFFFHFFFH',
    'FFFFFFFHFHFFHFFFFFFF',
    'HFHFFFFFFFFFFHFFFFFF',
    'HFFFFFFFFHHFHFFHHFFF',
    'FFFFFFFFFHFHFFFFFFFF',
    'FFFFFFFFFFFFHFFFFFFH',
    'FFFFFFFHFFFFFFFFFFFH',
    'FFFFFFHFFFFFFFFFHHFF',
    'HFFHFFFHHFHFFFHHFFFF',
    'FFFFFFFFFHFHFFHHHFFF',
    'HFFFFFHFFFFFHFHFFFFF',
    'HFFFFFFFFFFFFFFFHFFH',
    'FHFFFFFFFHFFFFFFFFFF',
    'FFHFFFFFFFHFFFFHFHFF',
    'FFHFHFFFFFFFHHFFFFFG'
]


class RTDP:
    def __init__(self, env):
        self.env = env
        self.iterations = 17000
        self.gamma = 1.
        self.goal_cost = -1.
        self.hole_cost = .1
        self.default_cost = .1

        self.actions = [0, 1, 2, 3]

        # FrozenLakeEnv gives us a transition matrix, the states and actions.
        # But no costs. So let's set something up ourselves.
        self._init_cost()
        self.calc_policy()

    def calc_policy(self):
        # RTDP
        V = defaultdict(lambda: 0.)

        for i in range(self.iterations):
            if i % 500 == 0:
                print(i)
            obs = self.env.reset()
            while True:
                action = np.argmin(
                    [
                        self.cost[obs][a] + self.gamma * sum(
                            item[0] * V[item[1]]
                            for item in self.env.P[obs][a]
                        )
                        for a in self.actions
                    ]
                )
                V[obs] = self.cost[obs][action] + self.gamma * sum(
                    item[0] * V[item[1]]
                    for item in self.env.P[obs][action]
                )
                obs, reward, done, _ = self.env.step(action)
                if done:
                    cost_ = -1
                    if reward == 0.0:
                        cost_ = 1
                    V[obs] = cost_ + self.gamma * sum(
                        item[0] * V[item[1]]
                        for item in self.env.P[obs][action]
                    )
                    break

        print(' ')
        print(V)
        self.V = V

    def _init_cost(self):
        # TODO Code below kinda ugly.
        cost = defaultdict(dict)
        for state in range(self.env.nS):
            for action in self.actions:
                # why 1? its the action itself. Others are slippery outcomes.
                if len(self.env.P[state][action]) == 1:
                    if self.env.P[state][action][0][2] == 1.0:
                        cost[state][action] = self.goal_cost
                    else:
                        cost[state][action] = self.hole_cost
                    continue

                done = self.env.P[state][action][1][3]
                if done and self.env.P[state][action][1][2] == 0.0:
                    cost[state][action] = self.hole_cost
                else:
                    cost[state][action] = self.default_cost

                if done and self.env.P[state][action][1][2] == 1.0:
                    cost[state][action] = self.goal_cost

        self.cost = cost

    def policy(self, state):
        return np.argmin(
            [
                self.cost[state][a] + self.gamma * sum(
                    item[0] * self.V[item[1]]
                    for item in self.env.P[state][a]
                )
                for a in self.actions
            ]
        )


def evaluate(rtdp, env):
    rewards = 0.
    state = env.reset()
    done = False
    steps = 0
    while not done:
        state, reward, done, _ = env.step(rtdp.policy(state))
        rewards += reward
        steps += 1
        if steps > 1e4:
            break

    return rewards


def run():
    env = FrozenLakeEnv(desc=MAP_20x20)
    rtdp = RTDP(env)
    tot_rewards = 0.
    eval_iter = int(1e4)
    for i in range(eval_iter):
        if i % 100 == 0:
            print(i)
        tot_rewards += evaluate(rtdp, env)
    print(tot_rewards / eval_iter)


if __name__ == '__main__':
    run()