results = pickle.load(f)
else:
results = Parallel(n_jobs=4)(delayed(onerun)(r)
for r in xrange(independent_runs))
with open("eu_prediction_sweep.pickle", 'wb') as f:
pickle.dump(results, f)
for r in xrange(len(results)):
avg_payouts[r, :] = results[r][0]
learned_actions[r, :] = results[r][1]
avg_payouts = avg_payouts.mean(axis=0)
learned_actions = learned_actions.mean(axis=0)
plot_that_pretty_rldm15(
[np.linspace(linspace_from, linspace_to, linspace_steps)],
[avg_payouts], ["EU"], "Prediction Accuracy", (0, 1.1, 0.2), "Payout",
(0, 1001000, 100000), 'eu_agent_payout.pdf')
plot_that_pretty_rldm15(
[np.linspace(linspace_from, linspace_to, linspace_steps)],
[learned_actions], ["EU"], "Prediction Accuracy", (0, 1.1, 0.2),
"Learned Action", (0, 1.1, 0.2), 'eu_agent_learned_action.pdf')
# fig = plt.figure()
# plt.xlabel('prediction accuracy')
# plt.ylabel('payout')
# plt.plot(np.linspace(linspace_from, linspace_to,
# linspace_steps), avg_payouts, label='EUAgent')
# plt.legend(loc='upper center')
# plt.savefig("eu_agent_payout.pdf")
# fig = plt.figure()
avg_payouts1.append(avg_payout1)
std_payouts1.append(std_payout1)
avg_payouts2.append(avg_payout2)
std_payouts2.append(std_payout2)
avg_total1.append(avg_totall1)
avg_total2.append(avg_totall2)
log.info('Average Payout: %.3f vs. %.3f (total: %.3f vs. %.3f)' %
(avg_payout1, avg_payout2, avg_totall1, avg_totall2))
avg_payouts1 = np.array(avg_payouts1)
std_payouts1 = np.array(std_payouts1)
avg_payouts2 = np.array(avg_payouts2)
std_payouts2 = np.array(std_payouts2)
avg_total1 = np.array(avg_total1)
avg_total2 = np.array(avg_total2)
plot_that_pretty_rldm15([
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payouts1, avg_payouts2], ["Defect", "Cooperate"],
"Cooperation Probability", (0, 1.1, 0.2), "Payout",
(0, 6, 1), 'defect_vs_cooperate.pdf')
plot_that_pretty_rldm15([
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_total1, avg_total2], ["Defect", "Cooperate"],
"Cooperation Probability", (0, 1.1, 0.2), "Payout",
(0, 7, 1), 'defect_vs_cooperate_total_payout.pdf')
avg_payout_sarsa = avg_payout_sarsa.mean(axis=0)
learned_actions_sarsa = learned_actions_sarsa.mean(axis=0)
avg_payout_eu = avg_payout_eu.mean(axis=0)
learned_actions_eu = learned_actions_eu.mean(axis=0)
avg_payout_avgq = avg_payout_avgq.mean(axis=0)
learned_actions_avgq = learned_actions_avgq.mean(axis=0)
y_range = (300000, 1001000, 100000)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payout_sarsa, avg_payout_avgq, avg_payout_eu],
["SARSA", "AVGQ", "EU"],
"Prediction Accuracy", (0, 1.1, 0.2),
"Payout",
y_range,
'figure_1_c_combined_newcomb_sarsa_avg_eu_payout.pdf',
custom_yticks=["%iK" % (int(x / 1000.0)) for x in np.arange(*y_range)],
fontsize=25,
label_fontsize=25,
label_offsets=[-30000, 0.0, 0])
y_range = (0, 1.1, 0.2)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [learned_actions_sarsa, learned_actions_avgq, learned_actions_eu],
[1001000, 1000]]))
agent1 = OneBoxNewcombAgent(problem1)
agent2 = TwoBoxNewcombAgent(problem2)
log.info('Playing ...')
log.info('%s' % (str(agent1)))
log.info('%s' % (str(problem1)))
log.info(' VERSUS')
log.info('%s' % (str(agent2)))
log.info('%s' % (str(problem2)))
_, payouts1 = interact_multiple(agent1, problem1, interactions)
_, payouts2 = interact_multiple(agent2, problem2, interactions)
avg_payout1 = payouts1.mean(axis=0)
avg_payout2 = payouts2.mean(axis=0)
avg_payouts1.append(avg_payout1)
avg_payouts2.append(avg_payout2)
log.info('Average Payout: %.3f vs. %.3f' % (avg_payout1, avg_payout2))
avg_payouts1 = np.array(avg_payouts1)
avg_payouts2 = np.array(avg_payouts2)
plot_that_pretty_rldm15([
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payouts1, avg_payouts2], ["TwoBoxer", "OneBoxer"],
"Prediction Accuracy", (0, 1.1, 0.2), "Payout",
(0, 1001001, 100000), 'one_vs_two_box.pdf')
avg_payout_eu[r, :] = euresults[r][0]
learned_actions_eu[r, :] = euresults[r][1]
for r in xrange(len(avgqresults)):
avg_payout_avgq[r, :] = avgqresults[r][0]
learned_actions_avgq[r, :] = avgqresults[r][1]
avg_payout_sarsa = avg_payout_sarsa.mean(axis=0)
learned_actions_sarsa = learned_actions_sarsa.mean(axis=0)
avg_payout_eu = avg_payout_eu.mean(axis=0)
learned_actions_eu = learned_actions_eu.mean(axis=0)
avg_payout_avgq = avg_payout_avgq.mean(axis=0)
learned_actions_avgq = learned_actions_avgq.mean(axis=0)
plot_that_pretty_rldm15([
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payout_sarsa, avg_payout_avgq, avg_payout_eu],
["SARSA", "AVGQ", "EU"], "Prediction Accuracy",
(0, 1.1, 0.2), "Payout", (0, 1001000, 100000),
'combined_newcomb_sarsa_avgq_eu_payout.pdf')
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [learned_actions_sarsa, learned_actions_avgq, learned_actions_eu],
["SARSA", "AVGQ", "EU"], "Prediction Accuracy", (0, 1.1, 0.2),
"Learned Action", (0, 1.1, 0.2),
'combined_newcomb_sarsa_avgq_eu_learned_action.pdf')
from plotting_stuff import plot_that_pretty_rldm15
if __name__ == '__main__':
from experiment_definitions import experiments
inputs = [e['picklefile'] for e in experiments]
if reduce(lambda a, b: a or b, map(lambda n: not os.path.isfile(n),
inputs)):
print >> sys.stderr, "run pd_two_player_generic.py first to \
create the .pickle files"
sys.exit(1)
for e in experiments:
with open(e['picklefile']) as f:
resultstruct = pickle.load(f)
results = resultstruct['results']
aparams1 = resultstruct['aparams1']
aparams2 = resultstruct['aparams2']
pparams = resultstruct['pparams']
expparams = resultstruct['expparams']
plotparams = resultstruct['plotparams']
for p in plotparams:
p['xdata'] = p['xdata'](results, e)
p['ydata'] = p['ydata'](results, e)
plot_that_pretty_rldm15(**p)
avg_payouts2.append(avg_payout2)
learned_actions1.append(learned_action1)
learned_actions2.append(learned_action2)
# TODO: this indexing is awkward, man why do you introduce 'metadata'
# above if you don't use it.
y_range = (0, 7, 1)
y_range_print = (1, 7, 1)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payouts1[0], avg_payouts2[0]], ["SARSA 1", "SARSA 2"],
r"$\epsilon$", (0, 1.1, 0.2),
"Payout",
y_range,
'pd_sarsa_avg_payout_normal.pdf',
custom_yticks=[""] +
["%i" % (int(x)) for x in np.arange(*y_range_print)],
fontsize=25,
label_fontsize=25,
y_lim=(0, 6),
label_offsets=[-0.1, -0.4, 0.0, 0.0])
y_range = (0, 1100000, 100000)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
],
[avg_payouts1[1], avg_payouts2[1]],
avg_payouts.append(avg_payout)
learned_actions.append(learned_action)
# TODO: this indexing is awkward, man why do you introduce 'metadata'
# above if you don't use it.
y_range = (0, 7, 1)
y_range_print = (1, 7, 1)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payouts[0], avg_payouts[1], avg_payouts[4], avg_payouts[5]],
["SARSA (I)", "SARSA (T)", "AVGQ (I)", "AVGQ (T)"],
"Prediction Accuracy", (0, 1.1, 0.2),
"Payout",
y_range,
'figure_2_c_combined_pd_sarsa_avg_payout_normal.pdf',
custom_yticks=[""] +
["%i" % (int(x)) for x in np.arange(*y_range_print)],
fontsize=25,
label_fontsize=25,
y_lim=(0, 6),
label_offsets=[-0.1, -0.4, 0.0, 0.0])
y_range = (0, 1100000, 100000)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps),
log.info(' VERSUS')
log.info('%s' % (str(agent2)))
log.info('%s' % (str(problem2)))
_, payouts1 = interact_multiple(agent1, problem1, interactions)
_, payouts2 = interact_multiple(agent2, problem2, interactions)
avg_payout1 = payouts1.mean(axis=0)
avg_payout2 = payouts2.mean(axis=0)
avg_payouts1.append(avg_payout1)
avg_payouts2.append(avg_payout2)
log.info('Average Payout: %.3f vs. %.3f' % (avg_payout1, avg_payout2))
avg_payouts1 = np.array(avg_payouts1)
avg_payouts2 = np.array(avg_payouts2)
y_range = (0, 1001000, 100000)
plot_that_pretty_rldm15(
[
np.linspace(linspace_from, linspace_to, linspace_steps),
np.linspace(linspace_from, linspace_to, linspace_steps)
], [avg_payouts1, avg_payouts2], ["TwoBoxer", "OneBoxer"],
"Prediction Accuracy", (0, 1.1, 0.2),
"Payout",
y_range,
'figure_1_a_one_vs_twoboxing.pdf',
custom_yticks=["%iK" % (int(x / 1000.0)) for x in np.arange(*y_range)],
fontsize=25,
label_fontsize=25)
