parameters_q_learning.py

Go to the documentation of this file.

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable

import os
import rospy

from rospkg import RosPack

# General parameters
ACTIONS = [(-0.6, 0.2), (0.6, 0.2), (0, 0.2)]
ACTION_COUNT = len(ACTIONS)

# Only use some of the LIDAR measurements
# When changing this value, also update laser_sample_count in q_learning.launch
LASER_SAMPLE_COUNT = 8


MODEL_FILENAME = os.path.join(RosPack().get_path(
    "reinforcement_learning"), "q_learning.to")

# Training parameters

# Start by loading previously trained parameters.
# If this is False, training will start from scratch
CONTINUE = False

DISCOUNT_FACTOR = 0.99  # aka gamma

MAX_EPISODE_LENGTH = 500
# Sample neural net update batch from the replay memory.
# It contains this many steps.
MEMORY_SIZE = 5000

BATCH_SIZE = 128
LEARNING_RATE = 0.0001

# Probability to select a random episode starts at EPS_START
# and reaches EPS_END once EPS_DECAY episodes are completed.
EPS_START = 1.0
EPS_END = 0.3
EPS_DECAY = 10000


class NeuralQEstimator(nn.Module):
    def __init__(self):
        super(NeuralQEstimator, self).__init__()
        self.fc1 = nn.Linear(LASER_SAMPLE_COUNT, 64)
        self.fc2 = nn.Linear(64, 32)
        self.fc3 = nn.Linear(32, ACTION_COUNT)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        return self.fc3(x)

    def load(self):
        self.load_state_dict(torch.load(MODEL_FILENAME))
        rospy.loginfo("Model parameters loaded.")

    def save(self):
        torch.save(self.state_dict(), MODEL_FILENAME)