Jump to content

  • Log In with Google      Sign In   
  • Create Account


Design of AI for my Game


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
8 replies to this topic

#1 Tourist   Members   -  Reputation: 108

Like
0Likes
Like

Posted 25 December 2013 - 08:16 AM

I am trying out a project for applying Neural Network in my game.

 

Game has a player who can move either up or down.

 

Objects: Coins which move in constant velocity towards the direction of a player and moves past if he doesnt catch

 

Coins generated in a random way originating from random spots 

 

Output: On whether the player should move up or down

 

Input: 

This is where am having troubles. Since the number of coins in screen and its position are going to be varying, i dont think that can be applied as input. I would like to know how to approach inputs in this ? Any similar articles where this is applied will help too. 

Thanks

 

 



Sponsor:

#2 Álvaro   Crossbones+   -  Reputation: 12494

Like
1Likes
Like

Posted 25 December 2013 - 11:38 AM

Perhaps you can divide the screen into some grid and use as inputs whether each grid is occupied or not, or perhaps how fast an object in the grid is going. You can make the grid have variable resolution, so the resolution is better near the player (the grid should probably be relative to the player).

Anyway, that's the first thing that came to mind.

#3 Tourist   Members   -  Reputation: 108

Like
0Likes
Like

Posted 27 December 2013 - 07:14 AM

Thanks . I will try that approach. 



#4 LorenzoGatti   Crossbones+   -  Reputation: 2621

Like
1Likes
Like

Posted 27 December 2013 - 08:00 AM

  1. If all coins are thrown towards the player, staying still is an optimal strategy that catches all coins. Can you explain coin movement more precisely? Does player motion affect coin movement? Also, how is the player's motion constrained? Looping movement (e.g. along a circle) would make the problem easier.
  2. Assuming that the purpose of the game is catching as many coins as possible and that some running around is needed, a straightforward neural network could be a bit too stateless: AI needs to plan what coins should be caught, and in what order, or let go. With n coins on the playfield, there are less than n!*2^n plans and a relatively long time to evaluate them.
  3. If you really need to use a neural network (homework?) an obvious brute force input choice is the number of coins that will cross each position of the player movement line at each future time (quantized into small bins), plus player position and velocity. You should be able to exploit and enforce translational symmetry along time and possibly space in network structure.

Produci, consuma, crepa

#5 IADaveMark   Moderators   -  Reputation: 2401

Like
1Likes
Like

Posted 27 December 2013 - 12:09 PM

Is the point of this exercise to create the best player or to use neural networks? Those 2 aren't necessarily the same.


Dave Mark - President and Lead Designer of Intrinsic Algorithm LLC

Professional consultant on game AI, mathematical modeling, simulation modeling
Co-advisor of the GDC AI Summit
Co-founder of the AI Game Programmers Guild
Author of the book, Behavioral Mathematics for Game AI

Blogs I write:
IA News - What's happening at IA | IA on AI - AI news and notes | Post-Play'em - Observations on AI of games I play

"Reducing the world to mathematical equations!"

#6 Tourist   Members   -  Reputation: 108

Like
0Likes
Like

Posted 27 December 2013 - 12:53 PM

@LorenzoGatti 
for the first point - answer is no. Coins are moving towards the direction of player and there is a vertical movement which will decide if a player will come in contact with the coin or not. 
 

for point 2  - easier decision is choosing which direction to move ? Would that be possible to design ? As am restricting directions to just up and down. (decision in my hands)

 

Point 3 - Tht looks like a gud suggestion too, but it looks like we are doing most of the work there :)

 

@IADaveMark 
Yes, I am attempting this project just to learn and get some practical experience on application of AI to Games. Am developing the game and trying to apply AI to the mainPlayer. ML approach need not necessarily be Neural Networks, but i thought its a decent choice. Correct me if am wrong there. 
 

Thanks 



#7 Álvaro   Crossbones+   -  Reputation: 12494

Like
0Likes
Like

Posted 27 December 2013 - 05:04 PM

You are probably wrong about neural networks. Neural networks have very limited applicability to games. You can actually compute where you have at what times to be able to catch each of the incoming coins, and then use some sort of dynamic programming algorithm to make a plan that maximizes the number of coins caught.

Also, how is this an application to games? About the only point of game AI for is to produce a good gaming experience for the player. For games like go this probably means making the AI agent as good as possible at playing the game. The situation you describe doesn't seem to have a player at all, or you are putting some algorithm in the place of the player; it isn't entirely clear to me why you would want to do that.

Game AI is about how the bad guys behave (and the occasional computer-controlled allied too).

#8 foxcode   Members   -  Reputation: 235

Like
0Likes
Like

Posted 28 December 2013 - 11:19 AM

Whilst I don't believe a Neural Network is appropriate for this, as Alvaro said, the applicability to games is limited and unnecessary when you can easily compute the correct answer, it can be done using a neural network.

 

Premise: A neural network usually has a fixed number of inputs. Since this game data in its raw form is variable, we need to do some pre-processing, sometimes called feature extraction. Pretend we are programming a neural network to identify letters in an image. The letter could be made from any number of pixels. To make the data fit the network we have to define features. A common one for this example would be dimension ratio. Letter height/width gives us a single value, and helps us identify the letter. A lower case 'o' for example would have a ratio of 1, where the letter 'l' may have a ratio of 2 as it's height is roughly double it's width.

 

You could use a similar system, one example feature could be the ratio of coins above the player to the ratio of coins below. Combine that with distance and direction of nearest coin, and you have a very simple network that will do what you ask. You could get more complex and calculate some fairly complex features, for example, total displacement to collect coins below against coins above.

 

To be honest the only reason I use neural networks for anything in games is personality. To change the behavior of a network, you simply load in a different set of weights and bam, you instantly have a more cautious/aggressive AI. But to sum up here is an example network for your game, it is very simple and could have many features added.

 

 

3 Inputs, 1 output

 

Input 1: Ratio of coins, number above/number below

Input 2: Distance player must travel to collect all coins above

Input 3: Distance player must travel to collect all coins below

Output 1: True means player travels down, false means player travels up

 

Hopefully something here will be helpful. In the above example, input 2 and 3 can be combined into a single input fairly easily. I think I would also have 2 outputs. This would allow you to pick a random direction when both outputs are false or true, to add some randomness to the AI.

Eg.

Output 1: True (Player should go up)

Output 2: True (Player should go down)    If you have a system like this, then you could pick randomly when ever the network produced conflicting results, for more precise play use a single output.



#9 IADaveMark   Moderators   -  Reputation: 2401

Like
0Likes
Like

Posted 30 December 2013 - 07:20 PM

To-date, I believe there are about 5 shipped titles that have used neural networks. And that's not for lack of interest in NNs... it's because of the lack of success in creating good game AI with NNs.


Dave Mark - President and Lead Designer of Intrinsic Algorithm LLC

Professional consultant on game AI, mathematical modeling, simulation modeling
Co-advisor of the GDC AI Summit
Co-founder of the AI Game Programmers Guild
Author of the book, Behavioral Mathematics for Game AI

Blogs I write:
IA News - What's happening at IA | IA on AI - AI news and notes | Post-Play'em - Observations on AI of games I play

"Reducing the world to mathematical equations!"




Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS