*Aqua*
-
Posts
1,119 -
Joined
-
Last visited
Content Type
Profiles
Forums
Developer Articles
KSP2 Release Notes
Posts posted by *Aqua*
-
-
1 hour ago, AngelLestat said:
But If you make the same ANN with many inputs from a microphone, a camera, and internet with another ann to understand arithmetic and language semantics. It would be able to learn that an elephant always come with 4 legs unless it is an injured elephant.
I have to correct you. You can only train an ANN to detect patterns. That's why it is useful for video and audio analyses because you are searching for patterns in there. However it can not for example extract new information from known one (one of the strong points of a semantic web) and neighter can it "think". It's essantially a set of functions, one for each output neuron. Therefore it can not "understand". There's no intelligence in there, the only intelligent thing is the trainer who forms the net as he wants it.
An ANN can only be a part of a hard AI. Self-awareness, "thinking", etc. must come from elsewhere.
-
I find this thread to be a good starting point. Also in this forum you should find the help you need.
There are several modders (although I can't name one) which create mods in Linux. I'm pretty sure you don't need Windows for that.
-
Is there a rule of thumb on how to archieve neutral buoyancy? Something like "for every ton you'll need 1000 intake liquid units"?
My first sub behaved weird. No matter how full the tanks were or wether the floaties from your survivability pack were inflated, the boat stayed at it's depth. I had to use the drive and the "wings" (don't know the word for it) to dive down or up.
The second boat just drowned like a stone because I left the lead in it. But how do I know how much lead I will need?And a suggestion:
To take the uboat to the water we either drive on wheels or fly it there (or cheat). It would be nice if there'll be a part which allows easy attachment of the boat to other parts like decouplers or docking ports.
Also, is there an end cap part planned? I first tried to build a boat without the inline propeller and was a bit sad that there's nothing to cover the end of the inline ballast and crew tanks.Edit:
-
10 hours ago, Speeding Mullet said:
Mk2 Docking port
I would like to have MK2 bicoupler & adapters in a LF-only variant. I really miss them.
Also a 2.5 m LF tank (rocket part) would be nice. I need it for bigger planes. -
-
There are two systems which render planets. When you are near a planet or moon the 3D model with heightmap will be rendered. From afar and in map mode/tracking station a simpler, scaled down variant will be used.
You made your screenshot just at the right time where the system switched between both modes.
-
Then you changed something. You need to post your log file.
-
But don't forget he eats potatos from Marsian soil. What kind of radiation is in there and how strong is it?
Afaik eating and inhalating radiated stuff is far more dangerous than just holding it in the hand.
-
You can also attach them straight and then rotate and translate them. Doesn't even take 10 seconds.
-
Because of the new heat system it would be useful to have a sensor part which reacts on a part's or vehicle's temperature.
Use case #1:
An automated probe fires retro rockets on decent if it gets too hot.Use case #2:
A plane autodeploys airbrakes depending on temperature. (We all know that flying >1.3 km/s in Kerbin's lower atmosphere will toast you.)Use case #3:
An interplanetary probe autoextends it's radiators when the nukes are nearly overheating.Use case #4:
As an extension of #3 it could be used to keep ISRU at the correct temperature for maximum efficiency. Also useful for other mods like KSP Interstellar where the nuclear reactors should run at specific temperatures. -
5 hours ago, AngelLestat said:
Mostly all the increase of breakthroughs in neuroscience are after the increase knowledge in ANN.
Which isn't surprising because an ANN is an abstract model of a real NN. ;-)
5 hours ago, AngelLestat said:About some of the things you mention on the elephant answer.. that can be solve just with normal ANN.
Nope. It can't really do that. Except you want an AI which makes errors.
ANNs are very good at patter recognition but that's it. For everything else there are much better and efficient algorithms. For example for learning behaviors an evolutionary algorithm learns much faster as an ANN and the results can be much more complex.
The idea is behind an evolutionary algorithm is that you take a piece of code or properties of an "entity" an modify them. Then you let it run wild and measure how good it performs. Then you modify it and let it run again. If it now performs better this "entity" will be chosen for the next generation. If it performs worse it'll be "killed". Now if you have several of them, each of them different, a near optimal "entity" will arise after several "generations".
An example for an evolutionary algorithm is here: http://math.hws.edu/eck/jsdemo/jsGeneticAlgorithm.html
The red things are supposed to find the green food. The more efficient they are at it the more likely it is they will survive. Just let it run for a while and you'll notice that they get better and better.This one is more visual. The goal is to drive as far as possible. Notice how the vehicles become more and more like cars.
I believe evolutionary algorithms will be the core of a hard AI.
-
18 hours ago, CliftonM said:
If you touch an outlet and get shocked, you're less likely to touch it again, as you now remember it as a danger. This also means that it would have to understand danger and self protection, along with self sacrifice, which would also come to the emotion part. Everything becomes a huge web of things that need to be done, and whenever a part is added, two more need to come along.
This "web of things" is already invented and it is called semantic web. The idea is that there are "concepts" (= things) that are set in relation to each other (= usually possession or identity). This simple idea can describe all kinds of information, no matter how abstract they are.
For example: Dumbo is Elephant. Elephant (has Leg) x 4. Elephant has (Color is Grey).
Question: Thing (has Leg) x4. Thing has (Color is Grey). Thing is ?If the system is programmed in a way that it assumes that it knows everything (closed world assumption) it will answer "Elephant".
If it is programmed in a way that it assumes it doesn't know everything (open world assumption) it will answer "Elephant OR unknown".Google search uses a semantic web to gather the information at the right of the search results list. I believe Google is the market leader in that technology atm.
8 hours ago, AngelLestat said:The new learning machines are not computers and they work the same way as our neurons works.
It may be nitpicking but that is actually wrong. Sure ANNs are modelled after NNs but they are only simplified models of the real thing. Organic neurons are pretty difficult to simulate because they are very complex.
In my computer science studies I also had several lectures about AI and intelligent systems. In my opinion for a hard AI we need four good designed systems:
1) perception (what do I sense?)
2) interpretation (what is it what I sense?)
3) evaluation (using the interpretations to make a description of what's going on)
4) planning (extrapolation of the current situation into the future, finding an appropiate reaction, defining a goal)Currently we can do 1) very good. 2) is so-so. 3) is impossible. 4) we still have no clue because we can't do 3).
Let me explain 3).
Evaluation here means that a system tries to group and categorize information. To use the example from before: It tries to understand a thing that is grey and has four legs. The problem is that there are numerous things that are grey and have four legs (cats, dogs, mice, etc.). The system can never be sure of what it is seeing. You can add more and more sensory data to make better guesses. But you need almost an infinite amount of data to be sure.
The next problem is that it has a lot of problems with new and contradicting data (an elephant with 3 legs?). The system can calculate a probability of how sure it is about a thing. But probabilities have the problem that you can't really rely on them. They are only a better way to guess and that's all.These problems persisted for decades and nobody found a reliable solution yet. And it doesn't look like someone will come up with something anytime soon.
-
Very nice!
I found a few quirks:
The scramjet has a spelling error. Or is it intended?
The OPT 2.5m-J Connector has 864 LF while the smaller OPT 2.5m-J Connector Variant has 1080 LF. The later one looks like it only has about 1/2 the volume of the former so it's weird. It's also weird that only so little fuel fits in there. The Rockomax X200-16 tank can hold about the same amount of fuel as the J Connector Variant but only has about 1/3 the volume of it.
-
46 minutes ago, RedParadize said:
Question is, does vertex/face normals affect how collider behave ?
No. Normals are only used for light calculations.
-
This test release doesn't include engines anymore? Or am I too stupid to find them?
-
I already guessed it'll be like that.
During the time I programmed a bit for KF I saw you jumping at every issue, bug, etc. This kind of affection coupled with a few programming difficulties and at the end being the only dev can cause a lot of stress. It was inevitable you would burn out.
Instead of giving up completely I propose you just pause until KSP 1.1 is out. The update will most likely break KF completely and I estimate it is still a few months away. When the update comes out how about rewriting KF together with me? We can also invite other people who are interested and then fix one feature at a time.
-
In the mean time you could try SmartParts. There's a proximity sensor in there (but it probably doesn't work the way you want).
-
Wildlife would be very nice.
But they should not have colliders.
Players are going to the extremes in this games which often means low altitude supersonic flight and (attempted) groundspeeds comparable to the maximum speed of racing cars. At those speeds you see the animals too late to evade them.And that's also a point. You don't see much of them when going fast. The physics bubble has a radius of 2.5 km around your craft. You need less than 10 seconds to cover that distance with mach 1. You probably only see them for just a few seconds before you whizzed past them.
-
It's a possible way but way to tedious to find my posts in a 100 pages thread.
-
6 hours ago, Temstar said:
You probably need to open up your case, take out the CPU cooler, unclip the cooling fan and vacuum the top of the heat sink.
Don't do that! A vacuum cleaner usually is highly electrostatic charged which can easily fry your computer! Instead remove the heat sink and then vacuum it or pull a tissue between the sinks. And don't forget to discharge yourself everytime before touching the insides of your computer.
-
For me sometimes it needs a bit of time after pressing enter before the embedded video shows up. Another user said it took 3 minutes on his computer.
-
Hello Kerbals!
How can I search for posts of a specific user in a specific thread?
Something like this will only return posts of someone calling me:
The advanced search doesn't have the ability to do what I'd like to do.
So... what do I need to do?
-
KSP on Linux always runs using OpenGL. DirectX is Windows-only.
-
Randomness isn't a good game mechanic if you don't give the player means to somehow counter it. That'll be difficult in KSP because usually almost every part of a rocket is critical for success. If there would be a failure somewhere it'll mean the rocket instantly turns into junk.
To be at the mercy of the RNGod is a bad thing.
[Plugin/Parts] Kerbal Foundries - Continuation [Latest: 1.9g]
in KSP1 Mod Development
Posted
I received your PM, lofi. Thank you! Now I know where send spam to.
I find your code isn't that unreadable. I recall I once sat down and dissected it. After two hours I was pretty sure I understood the important parts. That was right before I suggested a different class layout (if you remember).
I already send a pm to a KSP dev about 1.1 and he said that you can get the pre-release only 1-2 weeks before it's released. Sadly this isn't enough time to (probably) revamping everything on time.
I assume we (I?) have to start over with KF.