def main():
if len(sys.argv) > 1:
env = utils.setupEnv(sys.argv[1])
else:
env = utils.setupEnv(port=10000)
# Get the number of available states and actions - generates the output of CNN
observation_shape = env.observation_space.shape
action_size = env.action_space.n
#pdb.set_trace()
# Can start from a pre-built model
#load = input("Load model? y/n or an epsilon value to continue: ")
block_map_shape = (4,4,3)
myagent = agent(observation_shape, action_size,block_map_shape, True,0.0)
#pdb.set_trace()
scores = trainAgent(env, myagent)
'''
if load == 'y':
myagent = agent(observation_shape, action_size, block_map_shape,True,0.1)
#pdb.set_trace()
scores = testAgent(env,myagent)
elif load == 'n':
myagent = agent(observation_shape, action_size,block_map_shape)
#pdb.set_trace()
scores = trainAgent(env, myagent)
else:
#TODO - how come the 'epsilon value' runs still load a model??
myagent = agent(observation_shape, action_size, block_map_shape,True,float(load))
scores = trainAgent(env,myagent)
'''
np.savetxt('dqn_botscores',np.array(scores))
#plt.plot(scores)
#plt.show()
return
def main():
# Take in command line arguments, and use for environment setup
if len(sys.argv) > 1:
env = utils.setupEnv(sys.argv[1])
elif len(sys.argv) > 2:
env = utils.setupEnv(sys.argv[1], port=sys.argv[2])
else:
env = utils.setupEnv()
# Get the number of available actions
actionSize = env.action_space.n
# Give user decision on loading model or not
load = input("Load Q Table? y/n - Default as y:________")
# Set the Agent to Load Q-Table if user chooses to load
if load.lower() == 'n':
myAgent = QLearningAgent(actionSize,
200,
'QTable.json',
'qlearningResults.csv',
epsilon=1.0)
else:
myAgent = QLearningAgent(actionSize,
200,
'QTable.json',
'qlearningResults.csv',
True,
epsilon=1.0)
# Start the running of the Agent
myAgent.runAgent(env)
return
def main():
env = utils.setupEnv("MarLo-CliffWalking-v0")
# Get the number of available actions, minus waiting action
actionSize = env.action_space.n
epsilonDecay = 0.97
#not alpha
#have 6 values for gamma
# which ones are best then why we used values for the q-learning
#best two run for cliff and and treck a
gammas = [1, 0.8, 0.6, 0.4, 0.2, 0]
for gamma in gammas:
mc_QTableName = "mc_QTable_Gamma_" + str(gamma).replace(
".", "_") + "_Decay_" + str(epsilonDecay).replace(".",
"_") + ".json"
mc_CSVName = "mc_Results_Gamma_" + str(gamma).replace(
".", "_") + "_Decay_" + str(epsilonDecay).replace(".",
"_") + ".csv"
myAgent = MC_agent(actionSize, mc_QTableName, mc_CSVName, False,
epsilonDecay, gamma)
# Start the running of the Agent
myAgent.runAgent(env)
return
def main():
env = utils.setupEnv('MarLo-TrickyArena-v0')
# Get the number of available actions, minus waiting action
actionSize = env.action_space.n
epsilonDecay = 0.98
alphas = [0.8, 0.5, 0.1]
gammas = [1, 0.5]
for alpha in alphas:
for gamma in gammas:
QTableName = "QTable_Alpha_" + str(alpha).replace(
".", "_") + "_Gamma_" + str(gamma).replace(
".", "_") + "_Decay_" + str(epsilonDecay).replace(
".", "_") + ".json"
CSVName = "Results_Alpha_" + str(alpha).replace(
".", "_") + "_Gamma_" + str(gamma).replace(
".", "_") + "_Decay_" + str(epsilonDecay).replace(
".", "_") + ".csv"
myAgent = QLearningAgent(actionSize, 200, QTableName, CSVName,
False, epsilonDecay, alpha, gamma)
# Start the running of the Agent
myAgent.runAgent(env)
return
def main():
if len(sys.argv) > 1:
env = utils.setupEnv(sys.argv[1])
else:
env = utils.setupEnv()
# Get the number of available actions
actionSize = env.action_space.n
# Give user decision on loadind model or not
load = input("Load Q Table? y/n - Default as y:________")
# Set the Agent to Load Q-Table if user chooses
if load.lower() == 'n':
myAgent = MC_agent(actionSize)
else:
myAgent = MC_agent(actionSize, True)
# Start the running of the Agent
myAgent.runAgent(env)
return
def main():
# If arguments are supplied when running the agent, pass them to the setup env function, else use defaults
if len(sys.argv) > 1:
env = utils.setupEnv(sys.argv[1])
elif len(sys.argv) > 2:
env = utils.setupEnv(sys.argv[1], port=sys.argv[2])
else:
env = utils.setupEnv()
# Get the number of available states and actions - generates the output of CNN
observation_shape = env.observation_space.shape
action_size = env.action_space.n
# Initialise agent and then run it.
myagent = agent(observation_shape, action_size, False, 1.0)
scores = trainAgent(env, myagent)
'''
#Can start from a pre-built model
load = input("Load model? y/n or an epsilon value to continue: ")
if load == 'y':
myagent = agent(observation_shape, action_size, block_map_shape,True,0.1)
#pdb.set_trace()
scores = testAgent(env,myagent)
elif load == 'n':
myagent = agent(observation_shape, action_size,block_map_shape)
#pdb.set_trace()
scores = trainAgent(env, myagent)
else:
#TODO - how come the 'epsilon value' runs still load a model??
myagent = agent(observation_shape, action_size, block_map_shape,True,float(load))
scores = trainAgent(env,myagent)
'''
np.savetxt('dqn_botscores', np.array(scores))
#plt.plot(scores)
#plt.show()
return
def main():
env = utils.setupEnv('MarLo-CliffWalking-v0')
# Get the number of available actions, minus waiting action
#actionSize = env.action_space.n
actionSize = 5
epsilonDecay = 0.98
#alphas = [0.8,0.5,0.1]
#gammas = [1,0.5]
alphas = [0.8]
gammas = [1]
i = 1
for alpha in alphas:
for gamma in gammas:
QTableName = "QTable_Alpha_" + str(alpha).replace(
".", "_") + "_Gamma_" + str(gamma).replace(
".", "_") + "_Decay_" + str(epsilonDecay).replace(
".", "_") + ".json"
CSVName = str(i) + "_Test_Results_Alpha_" + str(alpha).replace(
".", "_") + "_Gamma_" + str(gamma).replace(
".", "_") + "_Decay_" + str(epsilonDecay).replace(
".", "_") + ".csv"
myAgent = QLearningAgent(actionSize,
25,
QTableName,
CSVName,
True,
epsilonDecay,
alpha,
gamma,
0.00,
training=True)
print(
"\n\n -------------- Starting test run of Decay %s, Alpha %s and Gamma %s --------- \n \n"
% (epsilonDecay, alpha, gamma))
# Start the running of the Agent
myAgent.runAgent(env)
return