tensorforce

example.py

import os
from collections import deque
import numpy as np
from tensorforce import Configuration
from tensorforce.agents.random_agent import RandomAgent
from tensorforce.agents import TRPOAgent
from tensorforce.core.networks import layered_network_builder


state = np.array([0] * 9, dtype='float')
actions = [
    ('foo',),
    ('bar',),
    ('baz',),
    ('bez',),
    ('boz',),
]


# random_config = Configuration(
#     states=dict(shape=state.shape, type=state.dtype),
#     actions=dict(continuous=False, num_actions=len(actions)),
# )
# random_agent = RandomAgent(random_config)
# a = random_agent.act(state)
# print(f'random action {a}')
# random_agent.observe(a, True)


trpo_config = Configuration(
    states=dict(shape=state.shape, type=state.dtype),
    actions=dict(continuous=False, num_actions=len(actions)),
    network=layered_network_builder([]),
)
trpo_agent = TRPOAgent(trpo_config)

model_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'trpo')
model_path = os.path.join(model_dir, 'model')
print(model_path)
if os.path.isdir(model_dir):
    trpo_agent.load_model(model_path)

rews = deque(maxlen=10)
for _ in range(10):
    a = trpo_agent.act(state)
    trpo_agent.observe(reward=abs(3 - a), terminal=True)
    rews.append(a)
    print('past reward {}'.format(sum(rews)))
trpo_agent.save_model(model_path)