def run_iters(self):
T = self.params['t']
results = {
'rewards': [],
'sup_rewards': [],
'surr_losses': [],
'sup_losses': [],
'sim_errs': []
}
trajs = []
beta = self.params['beta']
snapshots = []
for i in range(self.params['iters'][-1]):
print("\tIteration: " + str(i))
if i in self.params['update']:
self.lnr.train(verbose=True)
if i == 0:
states, i_actions, _, _ = statistics.collect_traj(
self.env, self.sup, T, False)
trajs.append((states, i_actions))
states, i_actions = utils.filter_data(self.params, states,
i_actions)
self.lnr.add_data(states, i_actions)
self.lnr.train()
else:
states, _, _, _ = statistics.collect_traj_beta(
self.env, self.sup, self.lnr, T, beta, False)
i_actions = [self.sup.intended_action(s) for s in states]
states, i_actions = utils.filter_data(self.params, states,
i_actions)
self.lnr.add_data(states, i_actions)
beta = beta * beta
if ((i + 1) in self.params['iters']):
snapshots.append((self.lnr.X[:], self.lnr.y[:]))
for j in range(len(snapshots)):
X, y = snapshots[j]
self.lnr.X, self.lnr.y = X, y
self.lnr.train(verbose=True)
print("\nData from snapshot: " + str(self.params['iters'][j]))
it_results = self.iteration_evaluation()
results['sup_rewards'].append(it_results['sup_reward_mean'])
results['rewards'].append(it_results['reward_mean'])
results['surr_losses'].append(it_results['surr_loss_mean'])
results['sup_losses'].append(it_results['sup_loss_mean'])
results['sim_errs'].append(it_results['sim_err_mean'])
for key in results.keys():
results[key] = np.array(results[key])
return results
def run_iters(self):
T = self.params['t']
results = {
'rewards': [],
'sup_rewards': [],
'surr_losses': [],
'sup_losses': [],
'sim_errs': []
}
trajs = []
d = self.params['d']
new_cov = np.random.normal(0, 1, (d, d))
new_cov = new_cov.T.dot(new_cov)
new_cov = new_cov / np.trace(new_cov) * self.params['trace']
self.sup = GaussianSupervisor(self.net_sup, new_cov)
snapshots = []
for i in range(self.params['iters'][-1]):
print("\tIteration: " + str(i))
states, i_actions, _, _ = statistics.collect_traj(
self.env, self.sup, T, False)
trajs.append((states, i_actions))
states, i_actions = utils.filter_data(self.params, states,
i_actions)
self.lnr.add_data(states, i_actions)
if ((i + 1) in self.params['iters']):
snapshots.append((self.lnr.X[:], self.lnr.y[:]))
for j in range(len(snapshots)):
X, y = snapshots[j]
self.lnr.X, self.lnr.y = X, y
self.lnr.train(verbose=True)
print("\nData from snapshot: " + str(self.params['iters'][j]))
it_results = self.iteration_evaluation()
results['sup_rewards'].append(it_results['sup_reward_mean'])
results['rewards'].append(it_results['reward_mean'])
results['surr_losses'].append(it_results['surr_loss_mean'])
results['sup_losses'].append(it_results['sup_loss_mean'])
results['sim_errs'].append(it_results['sim_err_mean'])
for key in results.keys():
results[key] = np.array(results[key])
return results