-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathdouble_q_learn_tabular_p_noise.py
123 lines (105 loc) · 3.74 KB
/
double_q_learn_tabular_p_noise.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
import itertools
import yaml
from collections import OrderedDict
num_seeds = 10
var_env_configs = OrderedDict(
{
"state_space_size": [8], # , 10, 12, 14] # [2**i for i in range(1,6)]
"action_space_size": [8], # 2, 4, 8, 16] # [2**i for i in range(1,6)]
"delay": [0], # + [2**i for i in range(4)],
"sequence_length": [1], # [i for i in range(1,4)],
"reward_density": [0.25], # np.linspace(0.0, 1.0, num=5)
"make_denser": [False],
"terminal_state_density": [0.25], # np.linspace(0.1, 1.0, num=5)
"transition_noise": [0, 0.01, 0.02, 0.10, 0.25],
"reward_noise": [0], # , 1, 5, 10, 25] # Std dev. of normal dist.
"dummy_seed": [i for i in range(num_seeds)],
}
)
var_configs = OrderedDict({"env": var_env_configs})
env_config = {
"env": "RLToy-v0",
"horizon": 100,
"env_config": {
"seed": 0, # seed
"state_space_type": "discrete",
"action_space_type": "discrete",
"generate_random_mdp": True,
"repeats_in_sequences": False,
"reward_scale": 1.0,
"completely_connected": True,
},
}
with open("tabular_rl/config.yaml", "r") as stream:
config = yaml.safe_load(stream)
env_name = config["env_name"]
agent_name = config["agent_name"]
agent_config = config["agents"][agent_name]
eval_eps = config["eval_eps"]
seed = config["seed"]
no_render = config["no_render"]
discount_factor = config["discount_factor"]
alpha = agent_config["alpha"]
episodes = agent_config["episodes"]
env_max_steps = agent_config["env_max_steps"]
agent_eps_decay = agent_config["agent_eps_decay"]
agent_eps = agent_config["agent_eps"]
# timesteps_per_iteration = agent_config["timesteps_per_iteration"]
agent_config = eval_eps
agent_config = {
# "env_max_steps": env_max_steps,
"num_episodes": episodes,
"epsilon_decay": agent_eps_decay,
"epsilon": agent_eps,
"render_eval": no_render,
"discount_factor": discount_factor,
"alpha": alpha,
"eval_every": eval_eps,
# "timesteps_per_iteration": timesteps_per_iteration, #todo: perhaps pass this later as an argument to the agent
}
algorithm = "double_q_learn_tabular_p_noise"
# agent_config = {
# "adam_epsilon": 1e-4,
# "beta_annealing_fraction": 1.0,
# "buffer_size": 1000000,
# "double_q": False,
# "dueling": False,
# "exploration_final_eps": 0.01,
# "exploration_fraction": 0.1,
# "final_prioritized_replay_beta": 1.0,
# "hiddens": None,
# "learning_starts": 1000,
# "lr": 1e-4, # "lr": grid_search([1e-2, 1e-4, 1e-6]),
# "n_step": 1,
# "noisy": False,
# "num_atoms": 1,
# "prioritized_replay": False,
# "prioritized_replay_alpha": 0.5,
# "sample_batch_size": 4,
# "schedule_max_timesteps": 20000,
# "target_network_update_freq": 800,
# "timesteps_per_iteration": 1000,
# "min_iter_time_s": 0,
# "train_batch_size": 32,
# }
model_config = {
"model": {
"fcnet_hiddens": [256, 256],
"custom_preprocessor": "ohe",
"custom_options": {}, # extra options to pass to your preprocessor
"fcnet_activation": "tanh",
"use_lstm": False,
"max_seq_len": 20,
"lstm_cell_size": 256,
"lstm_use_prev_action_reward": False,
},
}
value_tuples = []
for config_type, config_dict in var_configs.items():
for key in config_dict:
assert isinstance(
var_configs[config_type][key], list
), "var_config should be a dict of dicts with lists as the leaf values to allow each configuration option to take multiple possible values"
value_tuples.append(var_configs[config_type][key])
cartesian_product_configs = list(itertools.product(*value_tuples))
print("Total number of configs. to run:", len(cartesian_product_configs))