from VFA_Net import NeuralNetwork
nn_arq = [
{
"input_dim": 1,
"output_dim": 64,
"activation": "quadratic"
},
{
"input_dim": 64,
"output_dim": 1,
"activation": "none"
},
]
model = NeuralNetwork(nn_arq, bias=True, double=False)
A = np.array(1).reshape(1, 1)
B = np.array(1).reshape(1, 1)
Q = np.array(1).reshape(1, 1)
R = np.array(1).reshape(1, 1)
x0 = np.array(-1).reshape(1, 1)
u0 = np.array(0).reshape(1, 1)
# number of time steps to simulate
T = 30
# number of iterations of the dynamical systems for training
NUM_TRIALS = 250
ALPHA = 100
GAMMA = 0.9
"output_dim": 1,
"activation": "none"
},
]
ALPHA = 100
GAMMA = 1
NUM_TRIALS = 100000
BATCH_SIZE = 1
def loss(target, prediction, alpha=1):
return float((target - alpha * prediction)**2)
model = NeuralNetwork(nn_arq, bias=True, double=True, seed=1)
env = Blackjack()
# %%
def process(state):
"""
inputs a state, returns a parameterization of the state
uncomment return statements for different types of feature mappings,
don't forget to change the input dim of the neural net
"""
## normalized vector
ctc_arq = [
{
"input_dim": 3,
"output_dim": 512,
"activation": "relu"
},
{
"input_dim": 512,
"output_dim": 1,
"activation": "none"
},
]
ACTOR = PGNet(act_arq, bias=True, double=True)
CRITIC = NeuralNetwork(ctc_arq, bias=True, double=True)
ALPHA = 100
GAMMA = 1
NUM_TRIALS = 500000
BATCH_SIZE = 1
def loss(target, prediction, alpha=1):
return float((target - alpha * prediction)**2)
env = Blackjack()
# %%
import numpy as np
nn_arq = [
{"input_dim": 3, "output_dim": 512, "activation": "relu"},
{"input_dim": 512, "output_dim": 1, "activation": "none"},
]
ALPHA = 100
GAMMA = 1
NUM_TRIALS = 100000
BATCH_SIZE = 1
def loss(target, prediction, alpha=1):
return float((target-alpha*prediction)**2)
model = NeuralNetwork(nn_arq, bias=True, double=False, zero=True, seed=1)
env = Blackjack()
# %%
def process(state):
"""
inputs a state, returns a parameterization of the state
uncomment return statements for different types of feature mappings,
don't forget to change the input dim of the neural net
"""
## normalized vector
return np.array([state[0]/(max(env.state_space)[0]), state[1]/(max(env.state_space)[1]), state[2]]).reshape((3,1))
#TMAX = 20000 seems not needed? does the same thing as NUM_SAMPLES - I've replaced it in the implementation, you can always revert that
## slope and offset of function to be learnied
AA = 0.7
BB = 0.3
def f_star(x):
return AA * x + BB
def loss(target, prediction, alpha=1):
return float((1 / (alpha**2)) * np.square(target - alpha * prediction))
model = NeuralNetwork(nn_arq, bias=True, double=True, initVar=1, initVarLast=1)
def process(x):
return np.array(x).reshape(1, 1)
def sample_train(nsample):
xtrain = np.random.randn(nsample).reshape(nsample, 1)
return xtrain, f_star(xtrain) #xtrain, ytrain
def plot_loss(y):
fig, ax = plt.subplots()
label_fontsize = 18
from VFA_Net import NeuralNetwork
nn_arq = [
{
"input_dim": 1,
"output_dim": 64,
"activation": "quadratic"
},
{
"input_dim": 64,
"output_dim": 1,
"activation": "none"
},
]
model = NeuralNetwork(nn_arq, bias=False, double="yes")
A = np.array(1).reshape(1, 1)
B = np.array(1).reshape(1, 1)
Q = np.array(1).reshape(1, 1)
R1 = np.array(1).reshape(1, 1)
R2 = np.array(1.2).reshape(1, 1)
x0 = np.array(-1).reshape(1, 1)
u0 = np.array(0).reshape(1, 1)
# number of time steps to simulate
T = 30
# number of iterations of the dynamical systems for training
NUM_TRIALS = 1000
ALPHA = 100