class NNAgentSigmoid(Agent):
"""Template for neural network based agents"""
data_dir = Utility.get_data_path()
learning_rate = 0.01
def __init__(self, environment: Environment, seed: int, hidden_size: [int], activation_name: str = "relu"):
"""Load appropriate parameters depending on environment and learning time"""
super().__init__(environment, seed)
input_size = [self.environment.observation_length + self.environment.action_length]
output_size = [self.environment.reward_length]
self.nn = NeuralNetworkSigmoid(activation_name, input_size + hidden_size + output_size)
self.activations = ([], [])
def calculate_action(self, observation: str) \
-> str:
"""Feed percept into nn and calculate best activations action. Returns action."""
action = ""
reward = -2
# calculate expected reward by trying every action
number_of_actions = pow(2, self.environment.action_length)
if self.seeded_rand_range(0, 10) == 0:
action_idx = self.seeded_rand_range(0, number_of_actions)
action_string = format(action_idx, 'b').zfill(self.environment.action_length)
nn_input = NNUtility.bitstr_to_narray(observation + action_string)
nn_output = self.nn.forward(nn_input)
action = action_string
self.activations = nn_output
else:
for action_idx in range(number_of_actions):
action_string = format(action_idx, 'b').zfill(self.environment.action_length)
nn_input = NNUtility.bitstr_to_narray(observation + action_string)
nn_output = self.nn.forward(nn_input)
reward_string = NNUtility.narray_to_bitstr(nn_output[1][-1])
action_reward = Utility.get_reward_from_bitstring(reward_string)
if action_reward == reward:
i = self.seeded_rand_range()
if i == 1:
action = action_string
self.activations = nn_output
else:
if action_reward > reward:
action = action_string
reward = action_reward
self.activations = nn_output
return action
def train(self, reward: str):
"""Train agent on received reward"""
self.nn.backward(self.activations, NNUtility.bitstr_to_narray(reward))
import pickle
import matplotlib.pyplot as plt
from python.src import Utility
from matplotlib import rc
rc('font', **{'family': 'serif', 'serif': ['Computer Modern'], 'size': 12})
rc('text', usetex=True)
plt.rcParams['axes.axisbelow'] = True
data_dir_path = Utility.get_data_path()
plots_path = Utility.get_plots_path()
apiq_dict_path = data_dir_path.joinpath("apiq_dict.apiq")
apiq_dict = pickle.load(apiq_dict_path.open("rb"))
# bar plot of apiq values
f = plt.figure(figsize=(6, 4), dpi=400)
agents = list(apiq_dict.keys())
apiq_values = [v["mean"] for v in apiq_dict.values()]
apiq_errors = [v["error"] for v in apiq_dict.values()]
ypos = [len(agents) - 1 - y for y in range(len(agents))]
plt.barh(ypos[:4], apiq_values[:4], color="#076678")
plt.barh(ypos[4:5], apiq_values[4:5], color="#689d6a")
plt.barh(ypos[5:], apiq_values[5:], color="#8f3f71")
plt.errorbar(apiq_values,
ypos,
xerr=apiq_errors,
fmt=',',
ecolor='black',
capsize=4)
plt.yticks(
ypos,