def main(discount, n_objects, n_colours, n_trajectories, epochs, learning_rate,
structure):
# n_objects, n_colours 随便给的
"""
Run deep maximum entropy inverse reinforcement learning on the objectworld
MDP.
Plots the reward function.
grid_size: Grid size. int.
discount: MDP discount factor. float.
n_objects: Number of objects. int.
n_colours: Number of colours. int.
n_trajectories: Number of sampled trajectories. int.
epochs: Gradient descent iterations. int.
learning_rate: Gradient descent learning rate. float.
structure: Neural network structure. Tuple of hidden layer dimensions, e.g.,
() is no neural network (linear maximum entropy) and (3, 4) is two
hidden layers with dimensions 3 and 4.
"""
trajectory_length = 268
l1 = l2 = 0
env = Env(n_objects, n_colours)
# ground_r = np.array([env.reward_deep_maxent(s) for s in range(env.n_states)])
# policy = find_policy(env.n_states, env.n_actions, env.transition_probability,
# ground_r, discount, stochastic=False)
# trajectories = env.generate_trajectories(n_trajectories, trajectory_length,
policy = env.get_policy()
trajectories = env.generate_trajectories(2, trajectory_length,
lambda s: policy[s])
# feature_matrix = env.feature_matrix_deep_maxent(discrete=False)
feature_matrix = env.feature_matrix()
r = deep_maxent.irl((feature_matrix.shape[1], ) + structure,
feature_matrix,
env.n_actions,
discount,
env.transition_probability,
trajectories,
epochs,
learning_rate,
l1=l1,
l2=l2)
pkl.dump(r, open('deep_maxent_reward.pkl', 'wb'))
def train(grid_size, discount):
"""
Run linear programming inverse reinforcement learning on the gridworld MDP.
Plots the reward function.
grid_size: Grid size. int.
discount: MDP discount factor. float.
"""
env = Env(prepare_tp=True)
r = linear_irl.irl(env.n_states, env.n_actions, env.transition_probability,
env.get_policy(), discount, 1, 5)
pkl.dump(r, open("lp_reward.pkl", 'wb'))
r = maxent.irl(feature_matrix, env.n_actions, discount,
env.transition_probability, trajectories, epochs,
learning_rate, "alpha_%d.pkl", "alpha_205.pkl", 205)
pkl.dump(r, open("maxent_reward.pkl", 'wb'))
return r
if __name__ == '__main__':
train(0.01, 1, 400, 0.01)
rewards = pkl.load(open("maxent_reward.pkl", 'rb'))
env = Env(prepare_tp=True)
value = vi.value(env.get_policy(), env.n_states,
env.transition_probability, rewards, 0.3)
opt_value = vi.optimal_value(env.n_states, env.n_actions,
env.transition_probability, rewards, 0.3)
pkl.dump(value, open("maxent_value.pkl", 'wb'))
pkl.dump(opt_value, open("maxent_opt_value.pkl", 'wb'))
value = pkl.load(open("maxent_value.pkl", 'rb'))
opt_value = pkl.load(open("maxent_opt_value.pkl", 'rb'))
status = validate(value)
print(status)
pkl.dump(status, open("maxent_status.pkl", 'wb'))
status = validate(opt_value)
print(status)
pkl.dump(status, open("maxent_opt_status.pkl", 'wb'))