Gemini performance on FDCL/basic, broken down into groups

What this table contains:

Play mode – mStar = the number of move attempts until “mastery was demonstrated” (the bot makes 10 correct moves in a row). The smaller the number is, the faster the system learns. “Infinity” means that the run had to be interrupted (usually after ca. 300 move attempts) before such “mastery” could be demonstrated

Prepared-episodes mode: for how many future boards, out of 15, the bot could propose an error-free clearing plan. (15 = perfect result, 0 = never a success).

Rule set	Play-mode runs: m* (moves until mastery)	Prepared-episodes runs: Good future episodes (out of 15)
		Instructions w/o examples
Directory		prepared-1-b	prepared-2-b	prepared-3-b	prepared-5-b	prepared-10-b
Number of prepared episodes		1	2	3	5	10
Group 1: all pieces are movable. Static bucket assignment
cm_KRBY	22	10	15	10	10	9
cm_RBKY	20	10	15	10	7	15
quadMixed1	27	10	15	15	10	10
quadNearby	1	15	15	11	9	14
sm_csqt	10	0	15	2	10	0
sm_qcts	14	5	15	15	5	4
Group 2: all pieces are movable. Bucket determined by some sequence independent of game piece
buckets_2130	17	15	10	10	10	15
ccw	60	15	15	15	10	15
cw	37	15	15	15	15	10
cw_0123	6	15	10	2	12	10
Group 3: only some pieces are movable, according to some rule. Bucket does not matter
allOfColOrd_BRKY	39	2	0	0	0	0
allOfColOrd_KRBY	37	3	1	1	0	0
allOfShaOrd_csqt	105	0	0	0	0	0
allOfShaOrd_qcts	30	0	0	0	0	0
col1Ord_BRKY	Infinity	0	0	0	0	0
col1Ord_KRBY		0	0	0	0	0
colOrdL1_BRKY	Infinity	0	0	0	0	0
colOrdL1_KBYR		0	0	0	0	0
ordL1	16	0	0	0	0	0
ordRevOfL1	20	0	2	0	0	0
sha1Ord_csqt	77	0	0	0	0	0
sha1Ord_qcts		0	0	0	0	0
shaOrdL1_csqt	Infinity	0	0	0	0	0
shaOrdL1_qcts		0	0	0	0	0
Group 4: only some pieces are movable, according to some rule. Static bucket assignment
cm_RBKY_cw_0123	71	0	0	1	1	0
col1OrdBuck_BRKY0213	95	1	0	1	0	0
col1OrdBuck_BRKY3120		0	0	2	0	0
ordL1_Nearby	6	0	0	5	0	5
ordRevOfL1_Nearby	86	0	0	0	0	0
ordRevOfL1_Remotest	148	0	5	0	0	0
sha1OrdBuck_qcts0213	163	0	0	0	0	0
sha1OrdBuck_tqsc0213		0	0	0	1	0

Performance observations:

• Group 1: Gemini performs very well in both modes on “static” problems, where all game pieces are movable, and the destination buckets for game pieces are determined by a static map of some kind (which may be based on the game pieces’ properties and/or location, such as color match, shape match, or quadrant match).
• Group 2: Ditto for the Rules where all pieces are movable, while the destination is determined by some sequence that does not depend on the piece. (E.g. clockwise/counter-clockwise).
• Group 3: Only some pieces are movable, picked by some rule. The destination does not matter (e.g. ordL1). Here, Gemini in prepared-episodes mode fails, but the play mode often succeeds!
• Group 4: Only some pieces are movable, picked by some rule. The destination is determined by a static assignment rule (ordL1_Nearby).