artificial intelligence - Replacements for Neural Networks -

i learning artificial neural networks , usefulness intrigued me. tried make 1 play simple game (2048, done many times before seemed place start.)however, have found 2 problems along way.

the first way programming it, there no training data set. appears able worked around, haven't found way it.

the second neural networks, appear able minimize error, , game trying maximize score. @ given time there no optimal setup neural net check against.

i ability create ai , let learn best strategies without me directly teaching it, why decided try format in first place. question is, there type of nn can overcome these short fallings, or way program 1 without these issues, or have switch different algorithm/program. if have switch please give recommendation of switch to.

thank input

also, if belongs in different place in stack exchange please where

first, brief answer direct questions. read after recommendation :)

the first way programming it, there no training data set. appears able worked around, haven't found way it.

i don't think not having training data set @ outset hindering ability have training set. think generate own training set program plays game without nn. however, might not you're looking for.

the second neural networks, appear able minimize error, , game trying maximize score. @ given time there no optimal setup neural net check against.

this not problem, because define cost function make higher score have less error. perhaps can determine theoretical maximum score "0 error", , score of 0 error of 100, create cost function fits model.

what else there?

"reinforcement learning" seems fit closely 2048 in opinion, things ai won't know until try , have proof up! check out , see if it's you're looking for. brief quote link below.

reinforcement learning differs standard supervised learning in correct input/output pairs never presented...