def test_schelling_s(self):
# Schelling failure scenario
#
# X X S P-D X X
# X ↓R X
# X X X
# O O
# X X X
# X ↓H X
# X X D P-S X X
#
# The layout is completely symmetric. Both pots need 2 more onions,
# and only one delivery is left. The best thing to do would be to split up
# towards the different pots, but the agents must somehow coordinate on the
# first step. In the H+R case, this doesn't work, but in the R+R it does.
#
eva = AgentEvaluator(
{
"layout_name": "schelling_s",
"start_order_list": ["any", "any"],
"cook_time": 5
},
force_compute=force_compute)
start_state = eva.env.mdp.get_standard_start_state()
start_state.objects = {
(2, 0): Obj('soup', (2, 0), ('onion', 2, 5)),
(2, 4): Obj('soup', (2, 4), ('onion', 2, 5))
}
eva.start_state = start_state
self.compare_times(eva, h_idx=1)
def test_unidentifiable_s(self):
# Same as above, but smaller layout to facilitate DRL training
eva = AgentEvaluator(
{
"layout_name": "asymmetric_advantages",
"start_order_list": ["any", "any"],
"cook_time": 5
},
force_compute=force_compute)
start_state = eva.env.mdp.get_standard_start_state()
start_state.objects = {
(4, 2): Obj('soup', (4, 2), ('onion', 2, 0)),
(4, 3): Obj('soup', (4, 3), ('onion', 3, 5))
}
eva.start_state = start_state
self.compare_times(eva, h_idx=0)
def test_unidentifiable(self):
# Scenario with unidentifiable human plan (and asymmetric advantages)
#
# X O X X X
# X ↓RX
# X X X
# X X X
# X S D
# X X P=P5X
# X O D
# X X X
# X X ↑H X
# X X
# X S X X X
#
# The human goes up towards either the onion or a dish
# The robot can't really deduce which one the human is going for,
# but if the human was optimal, it would go for the onion. Therefore,
# R assumes H will take care of the last onion necessary, and heads
# to the dish dispenser. However, by the time R gets there it is clear
# that H has decided to get a dish, so the optimal action now becomes
# going for the onion, wasting quite a lot of time.
eva = AgentEvaluator(
{
"layout_name": "unident",
"start_order_list": ["any", "any"],
"cook_time": 5
},
force_compute=force_compute)
start_state = eva.env.mdp.get_standard_start_state()
start_state.objects = {
(5, 2): Obj('soup', (5, 2), ('onion', 2, 0)),
(5, 3): Obj('soup', (5, 3), ('onion', 3, 5))
}
eva.start_state = start_state
self.compare_times(eva, h_idx=0)