senorList = [item['state'] for item in state if item is not None]
nonFinalState = torch.tensor(senorList, dtype=torch.float32, device=device)
return nonFinalState, nonFinalMask
env = CooperativeSimpleMazeTwoD(config=config)
N_S = env.stateDim
N_A = env.nbActions
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [100]
netParameter['n_output'] = N_A
policyNet = MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
netParameter['n_output'])
targetNet = MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
netParameter['n_output'])
optimizers = optim.Adam(policyNet.parameters(), lr=config['learningRate'])
agent = DQNAgent(config, policyNet, targetNet, env, optimizers, torch.nn.MSELoss(reduction='none'), N_A, stateProcessor=stateProcessor)
agent.train()
nTraj = 100
config['logFlag'] = True
config['logFileName'] = 'SimpleMazeLog/DoubleQtraj' + mapName
config['logFrequency'] = 50
config['netUpdateOption'] = 'doubleQ'
env = SimpleMazeTwoD(mapName)
N_S = env.stateDim
N_A = env.nbActions
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [100]
netParameter['n_output'] = N_A
policyNet = MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
netParameter['n_output'])
print(policyNet.state_dict())
targetNet = deepcopy(policyNet)
optimizer = optim.Adam(policyNet.parameters(), lr=config['learningRate'])
agent = DQNAgent(policyNet,
targetNet,
env,
optimizer,
torch.nn.MSELoss(),
N_S,
N_A,
config['device'] = 'cpu'
# Get the environment and extract the number of actions.
# env = CartPoleEnvCustom()
trainEnv = gym.make("CartPole-v0")
testEnv = gym.make("CartPole-v0")
N_S = trainEnv.observation_space.shape[0]
N_A = trainEnv.action_space.n
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [40, 40]
netParameter['n_output'] = N_A
actorNet = MultiLayerNetLogSoftmax(netParameter['n_feature'],
netParameter['n_hidden'],
N_A)
criticNet = MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
1)
optimizer1 = optim.Adam(actorNet.parameters(), lr=config['learningRate'])
optimizer2 = optim.Adam(criticNet.parameters(), lr=config['learningRate'])
agent = ActorCriticTwoNet(actorNet, criticNet, [trainEnv, testEnv], [optimizer1, optimizer2], torch.nn.MSELoss(), N_A, config)
agent.train()
agent.test(100)
finalState = torch.tensor(sensorList, device=device, dtype=torch.float32)
return nonFinalState, nonFinalMask, finalState, finalMask
env = CooperativeSimpleMazeTwoD(config)
N_S = env.stateDim
N_A = env.nbActions
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [128]
netParameter['n_output'] = N_A
nPeriods = config['numStages']
policyNets = [MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
netParameter['n_output']) for _ in range(nPeriods)]
targetNets = [MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
netParameter['n_output']) for _ in range(nPeriods)]
optimizers = [optim.Adam(net.parameters(), lr=config['learningRate']) for net in policyNets]
agent = StackedDQNAgent(config, policyNets, targetNets, env, optimizers, torch.nn.MSELoss(reduction='none'), N_A, stateProcessor=stateProcessor)
policyFlag = True
if policyFlag:
for n in range(nPeriods):
return nonFinalState, nonFinalMask, finalState, finalMask
env = CooperativeSimpleMazeTwoD(config)
N_S = env.stateDim
N_A = env.nbActions
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [128]
netParameter['n_output'] = N_A
nPeriods = config['numStages']
policyNets = [
MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'], netParameter['n_output'])
for _ in range(nPeriods)
]
targetNets = [
MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'], netParameter['n_output'])
for _ in range(nPeriods)
]
optimizers = [
optim.Adam(net.parameters(), lr=config['learningRate'])
for net in policyNets
]
agent = StackedDQNAgent(config,
policyNets,
config['trainBatchSize'] = 32
config['gamma'] = 0.9
config['learningRate'] = 0.001
config['netGradClip'] = 1
config['logFlag'] = False
config['logFrequency'] = 100
config['priorityMemoryOption'] = False
config['netUpdateOption'] = 'doubleQ'
config['netUpdateFrequency'] = 1
config['priorityMemory_absErrUpper'] = 5
config['numWorkers'] = 4
env = StablizerOneD()
N_S = env.stateDim
N_A = env.nbActions
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [4]
netParameter['n_output'] = N_A
policyNet = MultiLayerNetRegression(netParameter['n_feature'],
netParameter['n_hidden'],
netParameter['n_output'])
optimizer = SharedAdam(policyNet.parameters(), lr=1.0)
agent = DQNA3CMaster(config, policyNet, env, optimizer,
torch.nn.MSELoss(reduction='none'), N_A)
agent.test_multiProcess()
config['mapHeight'] = 6
config['numAgents'] = 2
env = TwoAgentCooperativeTransport(config)
N_S = env.stateDim
N_A = env.nbActions
numAgents = env.numAgents
netParameter = dict()
netParameter['n_feature'] = N_S
netParameter['n_hidden'] = [128]
netParameter['n_output'] = N_A
policyNets = [
MultiLayerNetRegression(N_S[n], netParameter['n_hidden'], N_A[n])
for n in range(numAgents)
]
targetNets = [
MultiLayerNetRegression(N_S[n], netParameter['n_hidden'], N_A[n])
for n in range(numAgents)
]
optimizers = [
optim.Adam(net.parameters(), lr=config['learningRate'])
for net in policyNets
]
agent = MADQNAgent(config, policyNets, targetNets, env, optimizers,
torch.nn.MSELoss(reduction='none'), N_A)