-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun_rl_experiment.py
157 lines (149 loc) · 8.6 KB
/
run_rl_experiment.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
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
"""Script for running experiments in MDPs.
Using the parameter-based setting by default.
"""
import torch
import numpy as np
import random
import gym
import envs
import argparse
from mellog.logger import Mellogger
import utils
from algorithms.classic_thompson import GaussianDiscreteThompsonSampling
from algorithms.giro import GIRO
from algorithms.phe import PHE
from algorithms.optimist import OPTIMIST
from algorithms.randomistMCMC import RandomistMCMC
from algorithms.randomistMCMC_1step import RandomistMCMC_1Step
from algorithms.ucb1 import UCB1
from algorithms.gpucb import GPUCB
parser = argparse.ArgumentParser(description="Script for running a reinforcement learning experiment.")
# Parameters for any algorithm
parser.add_argument('--n_iterations', type=int, default=5000, help="Number of iterations")
parser.add_argument('--logging_freq', type=int, default=None, help="Logging frequency")
parser.add_argument('--random_seed', type=int, default=torch.randint(low=0, high=(2**32 - 1), size=(1,)).item(),
help="Random seed for the experiment")
parser.add_argument('--number_of_policies', type=int, default=49, help="Evaluation frequency")
parser.add_argument('--policy_type', default="linear", choices=["constant", "linear"], help="Policy class to be used")
parser.add_argument('--algorithm', default="randomist",
choices={"gaussian", "giro", "phe", "optimist", "ftl", "gpucb", "ucb1",
"randomist", "randomistMCMC", "randomistMCMC_1step"},
help="Algorithm to be run")
parser.add_argument('--env', default="car", choices={"car", "cartpole"}, help="Environment to be used")
parser.add_argument('--logdir', type=str, required=True, help="Name of the directory to log the results in")
parser.add_argument('--exp_name', type=str, required=True, help="Name of experiment. Should be the same for all runs")
parser.add_argument('--policy_std', type=float, default=0.15, help="Std for Gaussian policies (or hyperpolicies)")
parser.add_argument('--pseudo_rewards_per_timestep', type=float, default=1.1,
help="""Number of pseudorewards for unit of history in GIRO, PHE and RANDOMIST.
Give negative values for overriding with exploration with random returns.""")
parser.add_argument('--mcmc_steps', type=int, default=10, help="Number of MCMC steps in MCMC-RANDOMIST")
args = vars(parser.parse_args())
# Seed pseudo-random number generators
torch.manual_seed(args['random_seed']), np.random.seed(args['random_seed']), random.seed(args['random_seed'])
# Initialize logger
logger = Mellogger(log_dir=args['logdir'], exp_name=args['exp_name'], args=args, test_mode=False, dump_frequency=10)
# Initialize environment
if args['env'] == "car":
env = gym.make('MountainCarContinuous-v0')
parameter_range = torch.tensor([[-1, 1],
[0, 20]]).float()
return_range = [-5, 95]
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
horizon = 999
args['policy_std'] = torch.tensor([0.15, 3])
elif args['env'] == "cartpole":
env = gym.make('ContinuousCartpole-v0')
parameter_range = torch.tensor([[-2, 2],
[0, 4],
[0, 10],
[0, 12]]).float()
return_range = [0, 200]
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
horizon = 200
args['policy_std'] = torch.tensor([1., 1, 1, 1])
else:
raise ValueError
# Wrap the environment for pytorch use
env = utils.TorchTensorWrapper(env)
env.seed(args['random_seed'])
# Initialize algorithm
if args['algorithm'] == "gaussian":
algorithm = GaussianDiscreteThompsonSampling(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range,
policy_std=args['policy_std'], policy_type=args['policy_type'],
return_range=args['return_range'], setting="parameter_based")
elif args['algorithm'] == "giro":
algorithm = GIRO(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type=args['policy_type'],
pseudo_rewards_per_timestep=args['pseudo_rewards_per_timestep'],
setting="parameter_based")
elif args['algorithm'] == "phe":
algorithm = PHE(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type="constant",
pseudo_rewards_per_timestep=args['pseudo_rewards_per_timestep'],
horizon=horizon, mode='no_is', setting="parameter_based")
elif args['algorithm'] == "optimist":
algorithm = OPTIMIST(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type="constant",
horizon=horizon, trajectory_reuse=True,
ftl=False, setting="parameter_based")
elif args['algorithm'] == "ftl":
algorithm = OPTIMIST(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type="constant",
horizon=horizon, trajectory_reuse=True, ftl=True,
setting="parameter_based")
elif args['algorithm'] == "ucb1":
algorithm = UCB1(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type="constant",
horizon=1)
elif args['algorithm'] == "gpucb":
algorithm = GPUCB(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type="constant",
horizon=1)
elif args['algorithm'] == "randomist":
algorithm = PHE(state_dim=state_dim, action_dim=action_dim,
number_of_policies=args['number_of_policies'],
parameter_range=parameter_range, return_range=return_range,
policy_std=args['policy_std'], policy_type="constant",
pseudo_rewards_per_timestep=args['pseudo_rewards_per_timestep'],
horizon=horizon, mode='randomist', setting="parameter_based")
elif args['algorithm'] == "randomistMCMC":
algorithm = RandomistMCMC(state_dim=state_dim, action_dim=action_dim, parameter_range=parameter_range,
policy_std=args['policy_std'], alpha=2, eps=1, a=args['pseudo_rewards_per_timestep'],
nMCMCsteps=args['mcmc_steps'], return_range=return_range)
elif args['algorithm'] == "randomistMCMC_1step":
algorithm = RandomistMCMC_1Step(state_dim=state_dim, action_dim=action_dim, parameter_range=parameter_range,
policy_std=args['policy_std'], alpha=2, eps=1, a=args['pseudo_rewards_per_timestep'],
nMCMCsteps=args['mcmc_steps'], return_range=return_range)
else:
raise ValueError
total_return = 0
rets = []
for it in range(args['n_iterations']):
trajectory = utils.collect_trajectory(environment=env, agent=algorithm,
horizon=horizon)
# Add trajectory to buffer, triggering update of policy
algorithm.add_to_buffer(trajectory)
ret = float(torch.sum(torch.tensor(trajectory[2::3])).item())
total_return += ret
avg_return = total_return/(it+1)
logger.log("avg_return", avg_return)
logger.log("return", ret)
rets.append(ret)
if args['logging_freq'] is not None and it % args['logging_freq'] == 0:
print("[{}/{}] Return={:.2f}, Avg. Return={:.2f}".format(it, args['n_iterations'], ret, avg_return))