Skip to content

Commit 9b4c959

Browse files
Add Table IV verification
1 parent 036d6b4 commit 9b4c959

File tree

4 files changed

+578
-0
lines changed

4 files changed

+578
-0
lines changed

scripts/paper_results/confirm_paper_results.py

Lines changed: 68 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -46,6 +46,12 @@
4646
eval_sim_stack_cube = np.load(
4747
f"{paper_data_path}/TRI_SIM_STACK_CUBE.npy"
4848
) # Must be flipped for standard form
49+
eval_distribution_carrot_on_plate_supplement = np.load(
50+
f"{paper_data_path}/PU_HARDWARE_DISTRIBUTION_SUPPLEMENT.npy"
51+
) # Must be flipped for standard form
52+
eval_policy_carrot_on_plate_supplement = np.load(
53+
f"{paper_data_path}/PU_HARDWARE_POLICY_SUPPLEMENT.npy"
54+
) # ALREADY in standard form
4955

5056
# Load SAVI tests
5157
savi_hardware = MirroredSaviTest(alternative=Hypothesis.P0LessThanP1, alpha=0.05)
@@ -113,6 +119,68 @@
113119

114120
# Run appropriate tests on each data stream
115121

122+
####################################
123+
### Result 0: Princeton Hardware ###
124+
####################################
125+
carrotplate_distribution_result_lai_200 = lai_hardware_200.run_on_sequence(
126+
eval_distribution_carrot_on_plate_supplement[:, 0],
127+
eval_distribution_carrot_on_plate_supplement[:, 1],
128+
)
129+
carrotplate_distribution_result_step_200 = step_hardware_200.run_on_sequence(
130+
eval_distribution_carrot_on_plate_supplement[:, 0],
131+
eval_distribution_carrot_on_plate_supplement[:, 1],
132+
)
133+
carrotplate_distribution_result_savi = savi_hardware.run_on_sequence(
134+
eval_distribution_carrot_on_plate_supplement[:, 0],
135+
eval_distribution_carrot_on_plate_supplement[:, 1],
136+
)
137+
138+
carrotplate_policy_result_lai_200 = lai_hardware_200.run_on_sequence(
139+
eval_policy_carrot_on_plate_supplement[:, 0],
140+
eval_policy_carrot_on_plate_supplement[:, 1],
141+
)
142+
carrotplate_policy_result_step_200 = step_hardware_200.run_on_sequence(
143+
eval_policy_carrot_on_plate_supplement[:, 0],
144+
eval_policy_carrot_on_plate_supplement[:, 1],
145+
)
146+
carrotplate_policy_result_savi = savi_hardware.run_on_sequence(
147+
eval_policy_carrot_on_plate_supplement[:, 0],
148+
eval_policy_carrot_on_plate_supplement[:, 1],
149+
)
150+
151+
print()
152+
print("CARROT ON PLATE (Distribution Shift): ")
153+
print()
154+
print(
155+
"Lai-200 time-to-decision: ",
156+
carrotplate_distribution_result_lai_200.info["Time"],
157+
)
158+
print(
159+
"STEP-200 time-to-decision: ",
160+
carrotplate_distribution_result_step_200.info["Time"],
161+
)
162+
print(
163+
"SAVI time-to-decision: ",
164+
carrotplate_distribution_result_savi.info["result_for_alternative"].info[
165+
"Time"
166+
],
167+
)
168+
print()
169+
print("CARROT ON PLATE (Policy Shift): ")
170+
print()
171+
print(
172+
"Lai-200 time-to-decision: ",
173+
carrotplate_policy_result_lai_200.info["Time"],
174+
)
175+
print(
176+
"STEP-200 time-to-decision: ",
177+
carrotplate_policy_result_step_200.info["Time"],
178+
)
179+
print(
180+
"SAVI time-to-decision: ",
181+
carrotplate_policy_result_savi.info["result_for_alternative"].info["Time"],
182+
)
183+
116184
##############################
117185
### Result 1: FoldRedTowel ###
118186
##############################

0 commit comments

Comments
 (0)