def execute_maxent_causal(world, terminal, trajectories, discount=0.7):
"""
Maximum Causal Entropy Inverse Reinforcement Learning
"""
# set up features: we use one feature vector per state
features = world.state_features()
# choose our parameter initialization strategy:
# initialize parameters with constant
init = optimizer.Constant(1.0)
# choose our optimization strategy:
# we select exponentiated gradient descent with linear learning-rate decay
optim = optimizer.ExpSga(lr=optimizer.linear_decay(lr0=0.2))
# actually do some inverse reinforcement learning
reward = maxent.irl_causal(world.p_transition, features, terminal, trajectories, optim, init, discount)
return reward
def maxent(world, terminal, trajectories):
"""
Maximum Entropy Inverse Reinforcement Learning
"""
# set up features: we use one feature vector per state
features = W.state_features(world)
# choose our parameter initialization strategy:
# initialize parameters with constant
init = O.Constant(1.0)
# choose our optimization strategy:
# we select exponentiated gradient descent with linear learning-rate decay
optim = O.ExpSga(lr=O.linear_decay(lr0=0.2))
# actually do some inverse reinforcement learning
reward = M.irl(world.p_transition, features, terminal, trajectories, optim, init)
return reward
# re-compute detla for convergence check
delta = np.max(np.abs(omega_old - omega))
# re-compute per-state reward and return
return features.dot(omega)
# set-up the GridWorld Markov Decision Process
world, reward, terminal = setup_mdp()
# generate some "expert" trajectories (and its policy for visualization)
trajectories, expert_policy = generate_expert_trajectories(
world, reward, terminal)
# set up features: we use one feature vector per state
features = W.state_features(world)
# choose our parameter initialization strategy:
# initialize parameters with constant
init = O.Constant(1.0)
# choose our optimization strategy:
# we select exponentiated stochastic gradient descent with linear learning-rate decay
optim = O.ExpSga(lr=O.linear_decay(lr0=0.2))
# actually do some inverse reinforcement learning
reward_maxent = maxent_irl(world.p_transition, features, terminal,
trajectories, optim, init)
print("Done")
