fun_and_games/reinforce.py at master · sgould/fun_and_games · 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
149
150
151
152
153
# REINFORCE ALGORITHM ON GRID
# Stephen Gould <[email protected]>
#

import numpy as np
import matplotlib.pyplot as plt

import torch
import torch.nn as nn

import tqdm


class GridWorld:
    def __init__(self, n, m, init_pos, goal_pos):
        assert len(init_pos) == len(goal_pos) == 2
        assert (0 <= init_pos[0] < m) and (0 <= init_pos[1] < n)
        assert (0 <= goal_pos[0] < m) and (0 <= goal_pos[1] < n)

        self.n, self.m = n, m
        self.position = init_pos
        self.goal = goal_pos

    def move(self, direction):
        x, y = self.position
        if direction == 0:      # north
            self.position = (x, max(y - 1, 0))
        elif direction == 1:    # south
            self.position = (x, min(y + 1, self.n - 1))
        elif direction == 2:    # east
            self.position = (min(x + 1, self.m - 1), y)
        elif direction == 3:    # west
            self.position = (max(x - 1, 0), y)

    def at_goal(self):
        return self.position == self.goal

    def get_state(self, device):
        return torch.FloatTensor(np.concatenate((self.position, self.goal))).unsqueeze(0).to(device)

    def __str__(self):
        return "\n".join(["".join(["P" if self.position == (i, j) else "G" if self.goal == (i, j) else "." for i in range(self.m)]) for j in range(self.n)])


def sample_world(n=5, m=5):
    goal_pos = (np.random.randint(m), np.random.randint(n))
    #goal_pos = (0, 0)
    init_pos = (np.random.randint(m), np.random.randint(n))
    #init_pos = (m-1, n-1)
    while goal_pos == init_pos:
        init_pos = (np.random.randint(m), np.random.randint(n))

    return GridWorld(n, m, init_pos, goal_pos)


class PolicyModel(nn.Module):
    def __init__(self):
        super(PolicyModel, self).__init__()
        self.fc1 = nn.Linear(4, 16)
        self.fc2 = nn.Linear(16, 4)

    def forward(self, x):
        z = nn.functional.relu(self.fc1(x))
        y = nn.functional.softmax(self.fc2(z), dim=1)
        return y


def generate_episode(world, policy, device="cpu", max_episode_len=100):
    state = world.get_state(device)
    ep_length = 0
    while not world.at_goal():
        ep_length += 1
        p_action = policy(state).squeeze()
        log_p_action = torch.log(p_action)
        action = np.random.choice(np.arange(4), p=p_action.detach().cpu().numpy())

        world.move(action)
        next_state = world.get_state(device)
        reward = -0.1 if not world.at_goal() else 0.0

        sample = (state, action, reward)
        yield sample, log_p_action

        if reward == 0.0:
            break

        state = next_state
        if ep_length > max_episode_len:
            return

    sample = (world.get_state(device), None, 0.0)
    yield sample, log_p_action


def gradients_wrt_params(net, loss):
    for name, param in net.named_parameters():
        g = torch.autograd.grad(loss, param, retain_graph=True)[0]
        param.grad = g

def update_params(net, lr):
    for name, param in net.named_parameters():
        param.data += lr * param.grad


# --- tests ---

if False:
    world = sample_world()
    print(world, "\n")
    world.move(1)
    print(world, "\n")
    world.move(0)
    print(world, "\n")

    exit(0)

# --- learning ---

device = "cpu"
policy = PolicyModel()
policy.to(device)

lengths = []
rewards = []

gamma = 0.99
learning_rate = 1.0e-3

optimizer = torch.optim.AdamW(policy.parameters(), lr=learning_rate)
for episode_num in tqdm.tqdm(range(5000)):
    all_iterations = []
    all_log_probs = []
    world = sample_world()
    episode = list(generate_episode(world, policy, device=device))
    lengths.append(len(episode))
    loss = 0
    for t, ((state, action, reward), log_probs) in enumerate(episode[:-1]):
        gammas_vec = gamma ** (torch.arange(t+1, len(episode))-t-1)
        G = -0.1 * torch.sum(gammas_vec)
        rewards.append(G.item())
        policy_loss = log_probs[action]
        optimizer.zero_grad()
        gradients_wrt_params(policy, policy_loss)
        update_params(policy, learning_rate  * G * gamma**t)


plt.figure()
plt.plot(lengths)

plt.figure()
plt.plot(rewards)

plt.show()