train_policy = EpsilonGreedyPolicy(learning_algo, env.nActions(), rng, 1.)
test_policy = EpsilonGreedyPolicy(learning_algo, env.nActions(), rng, 0.1)
# --- Instantiate agent ---
agent = NeuralAgent(
env,
learning_algo,
parameters.replay_memory_size,
max(env.inputDimensions()[i][0] for i in range(len(env.inputDimensions()))),
parameters.batch_size,
rng,
train_policy=train_policy,
test_policy=test_policy)
# --- load saved network and test
# agent.setNetwork("test_4165747fe50541da92a5ea2698b190b90bc006d5.epoch=97")
agent.setNetwork(input_nnet) #tesot01
avg = agent._total_mode_reward
print(" _total_mode_reward: ", agent._total_mode_reward, ", nmbr of episode: ", agent._totalModeNbrEpisode, ", average per episode: ", avg)
Epoch_length = 500
mode = parameters.mode #mode 3 has planning depth 6#mode 2 ahs planning 3
agent.startMode(mode, Epoch_length)
agent.run(1, Epoch_length)
avg = agent._total_mode_reward / agent._totalModeNbrEpisode
print(" _total_mode_reward: ", agent._total_mode_reward, ", nmbr of episode: ", agent._totalModeNbrEpisode, ", average per episode: ", avg)
#just testing the saved nnet (possibly by visualizing the actions in the env)
show_score=True,
summarize_every=1))
# --- Run the experiment ---
try:
os.mkdir("params")
except Exception:
pass
# handle loading / saving weights
savedPath = fname + "_final"
if os.path.exists("nnets/" + savedPath): # ugly, but as in dumpNetwork
print("Loading saved net: " + savedPath)
agent.setNetwork(savedPath)
if test:
agent.startMode(PLE_env.VALIDATION_MODE, 10000)
dump(vars(parameters), "params/" + fname + ".jldump")
agent.run(parameters.epochs, parameters.steps_per_epoch)
if test:
agent.summarizeTestPerformance()
else:
# -- save network
agent.dumpNetwork(savedPath)
# --- Show results ---
basename = "scores/" + fname
scores = load(basename + "_scores.jldump")
plt.plot(range(1, len(scores['vs'])+1), scores['vs'], label="VS", color='b')
plt.legend()
