Reward gap model 20260405

[shnarazk]

"Conflict frequency" switches to "focus mode" if the average of conflict intervals is getting smaller. Therefore finding a "core" is the trigger. But it only supports learning-rate-based rewarding. Both picks recently used literals. We can't expect much.

"Reward gap model" is for the case we find new important-but-less-rewarded literals. Since they are new for conflict analysis, they would have small rewards and RL-based rewarding can't focus them immediately even the other literals have pretty big values (they are "core"). So switching to VMTF can handle the big range of rewards situations.

"splr",  1,"0a8a4c28d27228e954354ea0a6e7f16c-sum_of_three_cubes_42_kno",  TIMEOUT
"splr",  2,"1a936d41e3439d602c3ddcf96458a38c-arles_thres10_p10_r8185.c",    0.015
"splr",  3,"2b4467a5ac4ac41b36d4c3432b07f767-oddball_69_5_tto_zp.norma", 1921.757
"splr",  4,"3a1f1b4b9a521737cc760017fe9d8b43-MVRoundRobin_n16_d10_v3.c",  TIMEOUT
"splr",  5,"4ba2c1aa580b6497df6baf5e7e2c87be-at-least-two-vmpc_28.cnf" ,   64.577
"splr",  6,"5aed29ce52192a55ffbd2a6f340017e7-oisc-subrv-and-nested-12.",  TIMEOUT
"splr",  7,"6cb995b1c550beb579c53e27f6ca881a-RoundRobin_n16_d13.cnf"   ,  TIMEOUT
"splr",  8,"7a044c997ede14d00002f1db39d45170-sum_of_3_cubes_37_bits_87", 4821.444
"splr",  9,"8a05f9b6bf49285e40d0a197967ea5d3-arles_thres10_p10_r7466.c",    0.037
"splr", 10,"9a839badecb20dcf505ec79eedd3753a-anbul-dated-5-15-u.cnf"   ,  175.928
"splr", 11,"a0bcdaffb0ea36b678899fd86bdc7f18-arles_thres10_p10_r8186.c",    0.015
"splr", 12,"b1c8eaa002ac2fa1c8bfd1002738e78e-cliquecolouring_n15_k7_c6",  TIMEOUT
"splr", 13,"c0bd86bd7ca2c65e44311de374168150-goldcrest-and-14.cnf"     ,  TIMEOUT
"splr", 14,"d0298807e51730261ef65db827dcd70f-Break_triple_16_70.xml.cn", 1080.762
"splr", 15,"e2d2b011b0805782df6adba648db92e8-59-129706.cnf"            ,  926.741
"splr", 16,"f0bafebdcce23ccfbaf6c27a7522069b-div-mitern172.cnf"        ,  411.601
med:   293.764, max:  4821.444,total except 6 timeouts: 9402.877

Not so good. Reward itself is defined as an average. Its tendency is a second level average. And maybe it be difficult to define a meaningful threshold over varied problems.