def construct_policy():
training_dimensions = get_training_dimensions()
inputs, input_ends = get_nn_inputs(training_dimensions['state'])
action_inputs, action_ends = get_nn_inputs(training_dimensions['actions'])
x = tf.keras.layers.Concatenate()(input_ends)
x = add_dense_layers(128, 3, x, name='policy')
last_layer = x
outputs = {
name: method(last_layer, size, name)
for name, method, size in [
('which-action', create_categorical_output, Actions.NUM_ACTIONS),
('movement', create_linear_output, 3),
('collect-resources', create_categorical_output, Resources.NUM_RESOURCES),
('deposit-resources', create_categorical_output, Resources.NUM_RESOURCES),
('strafe-direction', create_categorical_output, 2),
('change-action', create_categorical_output, 2),
]
}
model = tf.keras.Model(inputs=inputs + action_inputs, outputs=[o for o in outputs.values()], name='ai_policy_function')
tf.keras.utils.plot_model(model, 'output/saved_networks/policy_function.png', rankdir='LR', show_shapes=True)
return model, [k for k in outputs.keys()], action_ends
def get_n_actions_input(actions, assignments):
num_games = len(actions)
training_dimensions = get_training_dimensions()
actions_input = initialize_n_input(training_dimensions['actions'],
num_games)
for idx in range(num_games):
actions_input['actions'][idx] = actions_to_numpy(
actions[idx], assignments[idx])['actions']
return actions_input
def get_policy_input_layers():
ret = {}
critic = PolicyHolder.get_policy()
training_dimensions = get_training_dimensions()
for input_key in training_dimensions['state']['inputs']:
for layer in critic.layers:
if layer.name == input_key:
ret[input_key] = layer
break
return ret
def get_critic_input_layers():
ret = {}
critic = SimpleCriticFunction.get_critic()
training_dimensions = get_training_dimensions()
for action_obj in ['state', 'actions']:
for input_key in training_dimensions[action_obj]['inputs']:
for layer in critic.layers:
if layer.name == 'flattened_' + input_key:
ret[input_key] = layer
break
return ret
def get_n_states_input(game_states, assignments=None):
num_games = len(game_states)
training_dimensions = get_training_dimensions()
state_input = initialize_n_input(training_dimensions['state'], num_games)
for game_idx, game_state in enumerate(game_states):
assignment = Assignment(
game_state['player-number']
) if assignments is None else assignments[game_idx]
game_encoding = game_state_to_numpy(game_state, assignment=assignment)
for input_key in training_dimensions['state']['inputs']:
state_input[input_key][game_idx] = game_encoding[input_key]
return state_input
def construct_critic():
training_dimensions = get_training_dimensions()
state_inputs, state_input_ends = get_nn_inputs(training_dimensions['state'])
action_inputs, action_input_ends = get_nn_inputs(training_dimensions['actions'])
x = tf.keras.layers.Concatenate()(state_input_ends + action_input_ends)
x = add_dense_layers(128, 3, x, name='critic')
x = tf.keras.layers.Dense(1, activation='sigmoid', name='the_output')(x)
model = tf.keras.Model(inputs=state_inputs + action_inputs, outputs=x, name='ai_critic_function')
tf.keras.utils.plot_model(model, 'output/saved_networks/critic_function.png', rankdir='LR', show_shapes=True)
return model
def construct_value_function():
training_dimensions = get_training_dimensions()
inputs, input_ends = get_nn_inputs(training_dimensions['state'])
x = tf.keras.layers.Concatenate()(input_ends)
x = add_dense_layers([1024, 512, 128], 1, x, name='value')
x = tf.keras.layers.Dense(1, activation='sigmoid')(x)
model = tf.keras.Model(inputs=inputs, outputs=x, name='ai_value_function')
tf.keras.utils.plot_model(model,
'output/saved_networks/value_function.png',
rankdir='LR',
show_shapes=True)
return model
def create_action_zeros():
training_dimensions = get_training_dimensions()['actions']
return {
input_key: np.zeros(training_dimensions[input_key])
for input_key in training_dimensions['inputs']
}