Stockfish 5 plays bad out of nowhere

I played a game against stockfish 5, in the opening and the begin of the middlegame he played well, but after move 22 he begins to play bad out of nowhere. Here you can see the game: lichess.org/study/Q2zHvsnN/UNoeSrNl#44

It is the biggest failing of engine programmers - they can make the algorithems stronger and stronger, but it is still beyond their capability to make engines play in anything that resembles a human style when set to low rating play- that is why the best they can do is hammer on you for 10-20 moves and then play a series of deliberately worse moves to try and even things up and if the position happens to have only one good move and 5-6 serious blunders then it will be forced to pick a blunder.

Best thing you can do with computers these days is to repeatedly force them to take back overly strong moves and make them play something more sensible looking , yes it is cheating, but they cheat on every single move they make so dont feel sorry for them.

I've done everything in my power to adjust Stockfish to try to play like a human, but even Google DeepMind can't solve that problem.

It's also worth noting that if this problem were easily solved, cheating would be commonplace and nobody would think they were playing against an engine.

Maybe you should try Shredder Chess: www.shredderchess.com/
It's considered to be one of the mos "human" chess software in terms of the problem mentioned above. There is also an app for mobile devices. Haven't try it yet so I can't confirm if it's true.
But of course it would be great if Stockfish could play more human-like as well. Or: Always play against Stockfih on the highest level and learn from your mistakes...;)

Honestly, what would probably best simulate human behavior would be a network of clusters of network of clusters of [etc.] of engines.

More affordably, perhaps run an ensemble of engines of varying strengths and let the policy node decide which engine's move gets played.

Could the playing strength of a machine be more uniformly diminished if the time allowed to make it's decisions was reduced, the memory available to it to make it's calculations was reduced and the maximum number of moves ahead it was allowed to calculate was reduced

+1 for Shredder, its playing style is very human-like (at least around 1600-1700 elo where I use it). It will occasionally make mistakes, but no more than humans do. Sometimes I even feel it deliberately overlooks moves that would be easy for an engine to see, but hard for a human.

The Windows version has a very easy to use UI, and it supports Chess960 very well. There's a free test/demo version you can download. The Android version (€7) uses the same engine and play is identical, but unfortunately there's no Chess960 support (and it's 4 years old now). Still the best app on Android I think.