nb_filters, nb_rows, nb_cols = 32, 3, 3
# keras model
keras_model = Sequential()
keras_model.add(Dense(args.hidden, activation="relu", input_dim=args.grid**2))
keras_model.add(Dense(args.hidden, activation="relu"))
keras_model.add(Dense(num_actions))
# X wrapper for Keras
model = KerasModel(keras_model)
# Memory
M = ExperienceReplay(memory_length=args.memory)
# Agent
A = DiscreteAgent(model, M)
# SGD optimizer + MSE cost + MAX policy = Q-learning as we know it
A.compile(optimizer=SGD(lr=0.2), loss="mse", policy_rule="max")
# To run an experiment, the Agent needs an Enviroment to iteract with
catcher = Catcher(grid_size=args.grid)
A.learn(catcher, epoch=args.epoch, batch_size=args.batch)
# Test the agent following the learned policy
A.play(catcher,
epoch=100,
visualize={
'filepath': args.output,
'n_frames': 270,
'gray': True
})
activation="softplus",
input_dim=agent_env.n_features))
keras_model.add(Dropout(0.5))
keras_model.add(Dense(hidden_dim, activation="softplus"))
keras_model.add(Dropout(0.5))
keras_model.add(Dense(hidden_dim, activation="softplus"))
keras_model.add(Dropout(0.5))
keras_model.add(Dense(agent_env.n_actions, activation="softmax"))
agent_model = KerasModel(keras_model)
# experience memory
agent_mem = ExperienceReplay(memory_length=memory_len)
# compile agent
agent = DiscreteAgent(agent_model,
agent_mem,
epsilon=lambda *args: epsilon)
# SGD optimizer + MSE cost + MAX policy = Q-learning as we know it
#agent.compile(optimizer=RMSprop(lr=0.001), loss='mse', policy_rule='max')
agent.compile(optimizer=RMSprop(lr=0.001),
loss='categorical_crossentropy',
policy_rule='max')
# train agent
agent.learn(agent_env, epoch=epochs, batch_size=batch_size, gamma=gamma)
# save trained model and weights
pre = "model-04-slow"
with open(pre + ".json", 'w') as f:
json.dump(keras_model.to_json(), f)
keras_model.save_weights(pre + ".h5", overwrite=True)