File size: 3,941 Bytes
93d8108
 
 
 
 
46b0409
8ae24a2
93d8108
 
46b0409
cf8b7c4
 
 
8ae24a2
93d8108
8ae24a2
 
cf8b7c4
 
 
93d8108
f929afb
 
93d8108
8ae24a2
93d8108
8ae24a2
93d8108
 
 
 
f929afb
 
 
 
cf8b7c4
 
 
46b0409
 
 
 
3266489
cf8b7c4
 
 
f929afb
 
 
 
93d8108
 
 
 
 
1af9e6d
 
93d8108
46b0409
f929afb
 
 
 
 
46b0409
 
 
 
 
cf8b7c4
 
 
 
8ae24a2
 
cf8b7c4
 
8ae24a2
f929afb
93d8108
 
 
786f010
17d4626
93d8108
cf8b7c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
93d8108
 
 
 
6a48762
93d8108
 
 
 
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
import gymnasium as gym
import numpy as np
from gymnasium.envs.toy_text.frozen_lake import generate_random_map
from matplotlib import pyplot as plt
from tqdm import trange
from AgentBase import AgentBase
import warnings


class DPAgent(AgentBase):
    def __init__(self, /, **kwargs):
        super().__init__(run_name=self.__class__.__name__, **kwargs)
        self.theta = kwargs.get("theta", 1e-10)
        print(self.theta)
        self.V = np.zeros(self.env.observation_space.n)
        self.Pi = np.zeros(self.env.observation_space.n, self.env.action_space.n)
        if self.gamma >= 1.0:
            warnings.warn(
                "DP will never converge with a gamma value =1.0. Try 0.99?", UserWarning
            )

    def policy(self, state):
        return self.Pi[state]

    def train(self, *args, **kwargs):
        i = 0
        print(self.gamma)
        while True:
            delta = 0
            V_prev = np.copy(self.V)
            for state in range(self.env.observation_space.n):
                # calculate the action-value for each possible action
                Q = np.zeros(self.env.action_space.n)
                for action in range(self.env.action_space.n):
                    expected_value = 0
                    for probability, next_state, reward, done in self.env.P[state][
                        action
                    ]:
                        if (
                            self.env_name == "CliffWalking-v0"
                            and state == self.env.observation_space.n - 1
                        ):
                            reward = 1
                        expected_value += probability * (
                            reward + self.gamma * self.V[next_state]
                        )
                    Q[action] = expected_value
                action, value = np.argmax(Q), np.max(Q)

                # update the state-value function
                self.V[state] = value
                delta = max(delta, abs(V_prev[state] - self.V[state]))
            if delta < self.theta:
                break
            i += 1
            # if i % 5 == 0 and i != 0:
            #     self.test(verbose=False)
            print(f"Iteration {i}: delta={delta}")

        self.Pi = np.empty((self.env.observation_space.n, self.env.action_space.n))
        for s in range(self.env.observation_space.n):
            for a in range(self.env.action_space.n):
                expected_value = 0
                for probability, next_state, reward, done in self.env.P[s][a]:
                    if (
                        self.env_name == "CliffWalking-v0"
                        and state == self.env.observation_space.n - 1
                    ):
                        reward = 1
                    expected_value += probability * (
                        reward + self.gamma * self.V[next_state]
                    )
                self.Pi[s, a] = expected_value
        idxs = np.argmax(self.Pi, axis=1)
        print(idxs)
        self.Pi = np.zeros((self.env.observation_space.n, self.env.action_space.n))
        self.Pi[np.arange(self.env.observation_space.n), idxs] = 1
        # print(self.Pi)
        # return self.V, self.Pi


if __name__ == "__main__":
    # env = gym.make('FrozenLake-v1', render_mode='human')
    dp = DPAgent(env="FrozenLake-v1", gamma=0.99)
    dp.train()
    dp.save_policy("dp_policy.npy")
    env = gym.make(
        "FrozenLake-v1",
        render_mode="human",
        is_slippery=False,
        desc=[
            "SFFFFFFF",
            "FFFFFFFH",
            "FFFHFFFF",
            "FFFFFHFF",
            "FFFHFFFF",
            "FHHFFFHF",
            "FHFFHFHF",
            "FFFHFFFG",
        ],
    )

    state, _ = env.reset()
    done = False
    while not done:
        action = dp.choose_action(dp.Pi, state)
        state, reward, done, _, _ = env.step(action)
        env.render()

    # plt.savefig(f"imgs/{0}.png")