The Big Programming Thread - Page 943

On February 03 2018 01:15 Silvanel wrote:
So i started learning C#. Not that i have other choice since company framework i need to use is exclusively written for C#. Any advices for beginner? (The only language i know already is Python).

from s in names 
     where s.Length == 5
     orderby s
     select s.ToUpper();

names.Where(s => s.Length == 5).OrderBy(s => s).Select(s => s.ToUpper());

On February 03 2018 01:15 Silvanel wrote:
So i started learning C#. Not that i have other choice since company framework i need to use is exclusively written for C#. Any advices for beginner? (The only language i know already is Python).

On February 03 2018 02:02 spinesheath wrote:

I don't know much python, so I can't really tell you stuff that would be relevant for someone switching from python to C#. Most important advice though: Get Resharper, preferably ultimate, and learn a new resharper command every day/week. 100% worth it. If your company doesn't already use it, try to convice them. Even if it only increases productivity by 1% it's probably well worth the cost already. It also makes refactoring easier, thus improving code quality and maintainability. And working in VisualStudio just is so much better with it than without.

Also get comfortable with the linq namespace. So this:

from s in names 
     where s.Length == 5
     orderby s
     select s.ToUpper();

and this:

names.Where(s => s.Length == 5).OrderBy(s => s).Select(s => s.ToUpper());

I prefer the latter version, though the difference is mostly aesthetical. Also Resharper can convert between the two versions with a few keypresses. Linq is one of the best parts of C# imo. To go with that, make sure you learn about and understand yield return and yield break.

C# is a nice language, but obviously not without issues. There are a few remnants of early versions of C# when it didn't support generics yet that still annoy me from time to time, but it rarely is a big deal. Also make sure you understand how events work with garbage collection, and read the msdn article on the disposable pattern. Otherwise you might run into a memory leak here and there.

On February 04 2018 02:34 mantequilla wrote:
is it possible to program a self learning AI in such a way that restricts developing of some behaviors?

like an AI that plays FIFA but doesn't use 1on1 tricks excessively. It would be lame if it tricks everyone and scores a goal that way.

On February 04 2018 03:42 Acrofales wrote:

Yes, it is.

Not very helpful answer, I admit, but machine learning relies on a reward (or cost) function. You can shape that any way you want. So if you can formalize the behaviour you want to exclude, you just penalize using it in the reward function.

On February 04 2018 18:36 emperorchampion wrote:


Until it teaches itself how to reprogram the reward...

On February 04 2018 19:14 Acrofales wrote:

Just in case you're serious (and if you are, stop watching Terminator and thinking that is how AI works): metalearning of that kind is currently not yet possible. But even if it was, it'd be something you, as the designer of the system, would control (at least enough to choose to implement it or not). So no. Current AI cannot teach itself to change the reward function.

On February 04 2018 20:01 emperorchampion wrote:


That’s exactly what an AI that could rewrite it self would say.

On February 05 2018 14:36 Mr. Wiggles wrote:
It's still sort of fun to observe what kind of unexpected behaviour can arise from a given reward function, even if you believe it should be limiting the behaviour of the agent.

My favourite example was from a grad student's demonstration in a reinforcement learning course I was taking. He had a Roomba that was using an optical sensor (or something like that) to tell if it was on a black 'track' in a confined rectangle. The reward function was for rotating the wheels forward while over the black track, while being punished for moving when off it or going in reverse. The rotation of the wheels was meant to be an approximation of movement. The idea was for the agent to eventually learn how to continuously move along the track.

In some situations however, the Roomba could accidentally wedge itself into a corner with it's wheels suspended while remaining over the track. It would then begin maximizing its reward by spinning the wheels at top speed while they were suspended.

The recent discussion reminded me of this, and it seemed like a funny anecdote about an agent finding an unexpected (and degenerate) way of maximizing a reward function.

On February 05 2018 07:26 mantequilla wrote:


only a true scheming AI would make such an accusation...

On February 05 2018 18:03 Silvanel wrote:
The vocabulary itself (reward function, agent, reinforcment) is causing us to project our emetions on it and humanize it. They should be using terms taken from psychology.

On February 08 2018 10:18 WarSame wrote:
My app has some DbAdapters and a Session to store the user's account.

Throughout my app I access the DB by creating a DB adapter, doing something with it, then closing it. I'm consider adding my DB adapters to either the Session or a separate object created at app startup to avoid having all that unnecessary code. Does this seem reasonable?

On February 08 2018 11:41 WarSame wrote:
Well I figure if I can do this I can remove a substantial amount of code. Instead of opening, using, then closing the adapter I can just have it open the whole time, and use it when I want it. This means that I can make inline calls to the DB without either wrapping it in a pointless function or cluttering up my code. I'm wondering whether my proposal is bad design, though.