aishenanigans

[aishenanigans]

Trying to teach a neural network to generate episodes of The Adventure Zone

So far, while my early attempts at neural nets have made some entertaining results, I haven’t played around with the code much to try to make them more or less accurate.

This time, I’ve taken the scripts of the first arc of The Adventure Zone: Balance, from tazscripts, and fed them to my neural net. Early results were not promising. It figured out their names, but that’s about it.

Clint: You shire somven; of me! Is fur the onwes fore you hear t-kinited uh, at a bith shat, us, soen ttank somad soinss, of rimns, to gic you… I ded that then spides doa gold for you attrl cqull Mwspereds—

Griffin: Uh—y, protige it eaving lits fory goff cithalide. I hill that— no’s coof over Travis Bhackin] Can’t that jusre.

Clint: Hemlinsid, inter, uhcerce that roke sialint mood wheresy  lion.

Griffin: Okay.

[laughon ahmit l

Griffin: ahene mery verly we’lk to talk the can’t as dooun a jue twing.

Griffin: Ehtue the lige on uldinged oighter— he see) Buld uh, that you cake twal's raughhis Coing thow it of reaale, gown’t glenoy of clidksee.

Clint: Har]

Justin: Dout vionee!

       Taako: I I goonde aocally bete] You weebans dighong that gutt is gons you wangher of diene] you naible noidhe— the morer, thor got the anksey if inded.

Travis: Bot, was flping.. Justin: pune.

Griffin: No, is fiesh telf that?

Griffin: Olly to have. Let thinhs, you sies, is wall] We ring it’s bars, this firing that’r thill and nath, lave be crom or this sponsting]

Griffin: It the sone.

Yeah so that wasn’t great. Let’s try turning down creativity.

First I tried changing “output bias”, but it wouldn’t run when I did that. Ok. Reset that back to the original value.

Then I tried changing the “hidden size” and “sequence length” hyperparameters. Didn’t seem to give me overtly better results. Monitoring the “loss” output, all tests started at at least 100% and leveled out at around 45-50%.

Changing the Wxh value from 0.01 to 0.1 made it forget how to spell “Griffin”.

Grrifirk voing. Om, axces ™he Ttay. A having. Okay. geh hore arpickighe tallen wale ot fot, and Banc a torterat nong om cave stuncleapmom,- thand thers Bhasping. Ot abe chack tore buwdy. It!

Grint: P/lavas whet you rint whas - thit s Youlko te wayn Yous centin? 1{'d be - a guwh Mau you tsiss toited, is the - lisesyot acke satite lit ta ymore. Ot Het urtleuht thith hire, crovas: So, you seet fomy you pagest. A go sy.

Ginfin: Is intauin re lont noup in? [tle jutclingtis a mort,

Thinking that I might be onto something, I changed that variable to 0.001. Still nonsense.

Ok, how about my Why value? A brief test with that at 0.1 (rather than 0.01) seemed to reduce loss after only a few thousand iterations, so I set Wxh, Whh and Why all to 0.1, and set it running. The loss was at 45% at around the 50000 iterations mark, which wasn’t notably better or worse than previous attempts. Left it running while I went to get my laundry out of the washing machine.

Even after 250000 iterations, I was still at 40% loss and nonsense came out. Further attempts necessary later.

Griffin: Uh, of, I cear that Dom it. [Clie. [taft Barnard?

Groen This in Briff sorne ase. Uhhino recklout to like forder uh, and you wall lork your!

KCliffinber I dof, it this omar and speres, thas it time -

Justin: Oht dest uh, and acty rich the a m—im theme]

Clin: Okay. Whet will prorto that

Sae it it Travis but in the is find it dimug: [lous: Ade Camed, uh! I can’t as cuning all you!  No, uh, uh,

Travis: Uh. An4. N5 what you gegh?

Griffin: Okayb onid, okay. And, s gist for op of it thoic Dlound is uh, you ya - yeah.

Griffin: [Taakus, intore. But, I'm -

Griffin: Uh, und carvers noc care vuctif the -

Griffin: [pare-

Cax towan Loundle ang is have the wald ton plack hive.

Magnus: Yea'

that chiog to on the goys a pree is. [lousthon ane up a wandy this, thas-in. Um oum, wourn, the stelliod yow that’s on you, uh, the uf allih my— it’s Yourd ants. Pensun the with, god - sors you s you

[aishenanigans]

First Explorations in Neural Networks - Featuring D&D Monsters

So, as a keen fan of this blog, and as someone working in the technology field (but unsure of what exactly I want my career to be), I decided it might be fun and interesting to build a neural network. It’s probably a useful, transferable skill, too.

For me, as a total beginner, the biggest stumbling blocks in getting this working were:

Figuring out what software I needed to get installed

Installing it, and getting it to work (I’ve just got a new PC with a ssd and hdd for the first time, and having a few issues with things installing in unexpected places...)

I installed, uninstalled, reinstalled a whole bunch. In the end, I’ve got Python 3 installed, along with about 5-7 other things (eg. scipy, numpy), and Anaconda, so that I can use the conda command prompt to run my code. I also installed Notepad++ in order to write my code - all the tutorials didn’t actually mention this, assuming I’d already have this basic level of knowledge! And I’m using regular Notepad to put my raw data in.

Initially, I followed a couple of tutorials, building a simple network using keras.

Once I was confident in my setup and how to input scripts, run them, and create datafiles, I found a very simple Recurrent Neural Network on GitHub which uses Python and numpy.

I got excited. I created a dataset (a list of D&D creature types), and I ran my script.

It failed.

Turned out this was written in Python 2, and I had Python 3.

Irritated at my noob mistake, but undaunted, I googled the error messages and edited the script until it worked. And the world was my oyster!

I cleaned up my initial dataset and ran my code for around 150000 iterations. This generated a bunch of D&D NPC names, demon/angel/deity names, names of creatures/monsters/races, and a bunch of words/names that really should never be used in a game by any sane individual! It enjoyed generating NSFW words; you can find those below the cut.

The Good

Krion (probably a young dragon)

Braender (elf)

Aryvir (elf)

Vodyitharnu (dragon)

Vorech (creature)

Shadran (morally dubious dwarf)

Shidlinor (innkeeper)

Uzicudache (roguish swashbuckler)

Uzifasleser (Uzicudache's younger brother)

Fion (tavern wench)

Vorgi'lle (orc)

Demora (some sort of demon woman)

Gara-sirgor

Aragir (lordly type)

Shadmorex (dracolich?)

Fricondrug (Viking type)

Ellit (bookish human)

Dracolax (a dragon that makes you poop)

Baladur

Gillon (probably a human blacksmith)

Parnau (rustic village in snow-covered mountains)

Barbasquex (idk but it sounds scary)

Gorgimoca

Nalkor beast

Pevitus

Parilith (she works in a temple of healing)

Grobizellan (sea captain)

Centifolk (sentient, intelligent centipedes)

Paphaspawn

Mezuma

Giant huspelver (bigger than your regular huspelver)

Fientyan Fig (wealthy, round man with a bald head)

Cengoin (dwarf whose father was Cengor)

Alane Marak (she owns a general goods store)

Caulin

Abbagon (lesser demon)

Beleriakes (elf city)

Crast Cracteroid (you don't trust him as far as you could throw him)

Phangoar

Deve Kromi (everyone knows someone like Deve)

Paviri (a deliciously-scented woman from a far-off land)

Omondworf (another Viking)

Bitor Wokolf (such Viking, wow)

Selanande Starticoria (sophisticated half-elf woman) 

Feendeller (bustling town)

Deptina (cute dwarf woman)

Nagmorithelf

Aarafolke

Fire of Fyandri (clearly a massively OP spell)

Grelia

Membrion Thue

Chymex

Threx

Xoroter

Skallira

Myrilith

Romoru

Spawn of Larntaner

Modrias Thod (he’s so dull that he makes ditchwater seem interesting)

Bent Conad (they call him bent because of his twisted leg)

Pegasta (you know her as Peggy)

Oodathi (consortium of exotic traders)

Fith Hassau

Orriel

Murkdrake

The Bad

Speder ol pigen (basically me, trying to talk about a fast pigeon, but drunk)

Nightie Seirite (has quite a lot of pizzazz to it!)

Mammath

Shocker

Al

Grey Rag

Blond

Calf

Ankel

Dinoman

Drakonic Dragon (as opposed to all the other sorts of dragon)

Hug

Winter

Goes Spere

Gray Desk (interesting that it came up with both “grey [noun]” and “gray [noun]”)

Gooda

Hammy

Cansaur

Horgie Seal

Wimin (ah yes, the more evil of the two races on this plane...)

Gnomz (gnomes are fine, but the youth gnomes of today, styling themselves as “gnomz”, are a scourge across the realms)

Half-gin

Rug

Mo

Sleet

Alan

Mommy

Boghester

Nick

Ouzu lord

Dan

Z (”My new character is only known as Z. That’s all he can remember. He remembers nothing of his past; that’s why I’ve not written a backstory.”)

Bard (the most evil obv)

Cool Byarnie (probably vapes)

Bee List (it’s like Bee Movie, but a list)

Gezuss (it tried so hard to spell “Jesus”)

NSFW

Bongher

Bongman

Peadman

Bundong

Phukman

Wagrider

Bonker

Ass

Tenis (unsure if this rhymes with “penis” or “tennis”)

Tit

Choder

As Beata

Wine Dong (my personal favourite)

Bummy

Fisst