Experimentations

Here are some notes.

• Selection of parameters
• Notes and choices for SDMS

Here is a set of experimentations.

• HSVI
• Q-Learning

BoxPushing

• HORIZON : 4

PBVI

Trial	Value	Size	Time
2	69,9301	214	28,5744

HSVI

HSVI	Num Trials	Error	Value LB	Value UB	Size LB	Size UB	Time	NumStates
Simultané	25	0,133875	98,5936	98,7275	13	29	>6000	80
Serial	8	0,03798	98,5936	98,6316	37	29	31,1217	45
Serial (tabular)	2	0,598185	98,5109	99,1091	13	13	5586,2	31871

HSVI Simultané

• LB : PLWC (type_selection=WCSP, pruning_freq=10, type_pruning=BOUNDED)
• UB : Sawtooth (type_selection=LP, pruning_freq=NO_PRUNING)
• HORIZON : 10

Problem	Num Trials	Error	Value LB	Value UB	Size LB	Size UB	Time	NumStates
alignment	73	43,5217	91	134,522	50	526	7083,55	1033
mabc1	1	0	9,29	9,29	20	10	0,815401
boxPushing	2	13,4805	215,23	228,711	28	18	6007,43	38
Mars	1	2,66062	25,9527	28,6133	20	10	4399,63	20
recycling	12	0	31,8639	31,8639	15	71	14,8543	104
tiger	21	26,8749	3,90758	30,7825	30	138	6148,64	203