WORKAROUND: ObHack-6-Determinisitc is still not 100% deterministic

[samboy]

While ObHack-6-Deterministic is almost entirely deterministic, MAP06 when using the seed 1FreeDoom1 and the default parameters still varies between invocations of ObHack. All of the other maps are the same, but there still is some lingering non-deterministic behavior, even though I have stopped using pairs() in the Lua code.

[samboy]

Related: https://github.com/samboy/ObHack/issues/4 We’re doing a lot better than we did before.

[samboy]

After a lot of trouble, I thought I found it. In planner.lua, on line 2732:

if a.combo == b.combo then return false end

Here a.combo and b.combo are both tables. However, tables are referenced by pointer, not by value, so this equality test merely asks “are these tables the exact same tables?”, not “do these tables have the same values?”

I’m not sure what the fix will be. Probably comparing values in the tables instead of just comparing the table pointers.

[samboy]

Actually, this is the line which sometimes (but not always) varies in planner.lua:

if a.f_min < (b.f_max + 64) then return false end

Got me. I can’t readily reproduce it; other bits of code also slightly randomly non-deterministically vary. Ugh.

[samboy]

It’s b.f_max which varies here.

[samboy]

It varies because of this code in planner.lua in the function minmax()

if link.build ~= c then

Like I discuss above, this is the error of comparing a table pointer instead of a table.

[samboy]

Here is the workaround I will do so that 1FreeDoom1 and some other wads are deterministic:

        -- Ugly hack so "1FreeDoom1" always makes the same megawad
        -- There is a bug which I will not track down (right now) that
        -- makes MAP06 with the seed "1FreeDoom1" somewhat vary.
        -- Finding the variance is too much work, so I will instead just
        -- make a different MAP06.
        -- See https://github.com/samboy/ObHack/issues/8
        if index == 6 then con.random() end

I hate closing the bug without a real fix, but this is “good enough” for now. I might reopen this in the future if I ever feel like jumping down this rabbit hole again.

[samboy]

Still varies. Will look at tomorrow.

[samboy]

The variations appear to come from glBSP, which can make different .wad files when given the same map as source. If I run this seven times:

./ObHack 1FreeDoom1

I get the exact same (same SHA-256 sum) file seven times in a row. Ditto when the seed is 2FreeDoom1, so I am closing this issue.

It’s just too deep of a rabbit hole for me to try and make this old ObHack code 100% deterministic unless I get really board — not my life these days as a single parent.

[samboy]

While the .wad files (even the non-glBSP .wad files) differ, if I change line 952 of writer.lua to read if true (instead of if false), so that it makes .txt files (against the 2020-02-11 tag/release), I get the following SHA-256 sums when making the MAPXX.txt files using the seed 1FreeDoom1:

426e6beb16bdd762e85b9ab4afb751977ced1fc3a5fde3e4170df6c85c3388ff *MAP01.txt
3d5b6d1238005f38a78ed8014ff19afc9790e991987ebdada5b0db72ec9a4c83 *MAP02.txt
271b6bd78830b272d85998a56ab3944f5b6356496729c1759e5e6ab617fbdeb7 *MAP03.txt
d90554647c38410464e4a7c85b03f638965a73a98a1dadc1a94dbaf9b04372ed *MAP04.txt
ba0dc8d6586c4e2d830053e58dd38a3a36607759d163ec55cb066536fc65980b *MAP05.txt
b5038822194623c5d977964c261d1504100c6fcfca85cf9ae6b797d3488ce35e *MAP06.txt
2f59b08b714260831404dbac912ecf1cf3a3a6a862ac11bd3835bb31a237571a *MAP07.txt
2c7379777c30857809ec52dc00654b27d2079fc26cc19346a38560f66f61d9cf *MAP08.txt
c280ad6988f41bfb2476ccc9749c12903bb524bcafd119f5b52ad0b350fc4ed6 *MAP09.txt
4c9498bfb5f353394a52926c5fa80b2ade13f0cd0dc668e0bf3c920befced29e *MAP10.txt
98c9fe685574e037d1bbb8f1d2eea802e4222aa5aa8fb020008f0bec4c90aaa8 *MAP11.txt
d709a8525f2c968def4955bf75164a2ececd84397dcb546cc30005ba76e515ac *MAP12.txt
ae436ce1f8f7ed8137ffcb1a3a24dd032752af47ed6cc3db7b387e4651272af6 *MAP13.txt
9beb84ced413e4f538bd081faeafd02daecf3197c7ee37d511bc144d4a82d6fe *MAP14.txt
f05bd1fe38e79250fa5c18db935804665c3359c3ed9d7b52731494040cf4fc05 *MAP15.txt
86f4165b29e6530ac55bd75a36f2720ae1c79d4f61547bd78c3550bd3ea757d6 *MAP16.txt
e172246cb974b5fe9fb4355d54450289c14f670bfe2e24ce91a90ace185e5c1f *MAP17.txt
f68dc962c72f500f79c70f710435b8d157ea1d78b082bd84b329e05a3e74019e *MAP18.txt
d9518d79f07a2e3b9b885e4a321f18d0beddf8f28bfff565298e5277edb825e5 *MAP19.txt
4edef90eaecc3aba9abbff48598210933b70445939c34d0346d8b89fa4ea1510 *MAP20.txt
85bb93a61b4b91d2388ac4cd167db7a206a8fa6ad5cadc3df55a178fa7aba023 *MAP21.txt
ee44c09351abb6881e4e064d0657b8fc3ef79b74e33aae84b12bcf4c439ad35b *MAP22.txt
bcaa4ccac63816166a969eb419e4ff7a0eaf7a865261de06ae5b0b694f7e8c35 *MAP23.txt
b3fd4c68ad402bc7192f9f002065178e686356dd5897e073f6771ca1701709ce *MAP24.txt
8f3a1865739042986277641735670b9def6bdf039bcd79173bc64f36eb0e4f75 *MAP25.txt
6f292cb794ee84f6832a24533a659d0be4d18f81ac127897d2bff4b63f90d414 *MAP26.txt
a93a7b01c79170dc9f9c471f4179ce99b4241301e9c9f9bc377005d59add68fb *MAP27.txt
b1b57f5af4affb619fdb0d31ebe31447e1f33d0da0e1db5258fd38fddbec262d *MAP28.txt
5d2864ff153d9267e1beb55587eb7a43741bbd4575a6995d7e010190cd296ab0 *MAP29.txt
aa92847d525e7e52f77a8af0a1dbc30a2bd84c1c07f2689cbb832a26e7b6892a *MAP30.txt
34ffd81843412c2b7582195cd166f903451fc30f63d847dd59148e0888096909 *MAP31.txt
58c52624043390cf8d492393096f7930484e938518a9c9695dcc2333410d39b8 *MAP32.txt

[samboy]

However, TEMP.wad can differ, as can the glBSP-generated 1FreeDoom1.wad files, so there is some non-determinism in writer.lua. But since the maps themselves are the same (even when the .wad files have variation), this is not important enough of an issue right now. Back to MaraDNS.