Fighting Game AI
Author
I'm making Mortal Kombat vs Street Fighter game, and was wonder if anyone knows how to implement AI for fighting games. It would help me out a lot. Thanks.
The simplest would be a weighted system, basically fighting sprite guy makes a list of all possible moves and then goes over several rules that add or subtract from the weight, like if a option hurt the opponent +5, if it leaves the sprite vulnerable -1, if it keep it from getting hurt +3. Then take some of the ones with the higher weight (better choices) and randomly chooses one, this step makes it so you sprite wont always choose the best choice but still a good one so its unpredictable but not random. With this you can adjust how much weight is put on each rule to choose style and how smart or random your sprite will be.
A neural network is another option; it would be more complex but could be interesting.
A neural network is another option; it would be more complex but could be interesting.
Quote:Original post by Kaze
A neural network is another option; it would be more complex but could be interesting.
It shouldn't be too difficult, since a 2d fighting game has complete information, no randomness, very little planning and above all very cheap, meaningful and tractable AI inputs and outputs.
Inputs can be (for each character) 2D position, pose/state (e.g crouching), hit points (important to choose a final blow) and "energy" meters;outputs are a small set of discrete moves (possibly the same keypress commands the player uses): jump up, leap forward, turn, walk, dash, various attacks...
The NN can be blindly evolved, with no assumption other than left/right symmetry: in matches between humans their actions can be recorded (this feature would also useful for action replays) and used as training objectives; ensembles of initially random networks can be trained by fighting between each other and against humans; the AI for one character can be used as a starting point for the AI of another.
Different strengths of AI can be obtained by taking stronger or weaker specimens from a population of different neural networks.
Lorenzo Gatti
Firstly you should have a good FSM system in place for your fighting system. You could use one of the free FSM generaters available on the internet. The FSM generater also exports the entire FSM state machine in JAVA/C++ code. Once this system is set, then you must start with your AI part. The AI can be broken into different states , action, style, balance. When the AI is set to action state, it will see to it that the AI is always in attack mode. When the AI is set to style state, it will see to it that the AI is representing the characteristics of the character. So character specific styles can be seen. When the AI is set to balance state, it will see to it that the AI is balances the other two states. This will see to it that the AI is not too aggressive nor too defensive.
Also a complete weighting system must also be in place with regards to scoring. One could divide the characters into different parameters like agility, strength, charisma, etc. One could then use this for calculating the hurt scoring mechanism. I hope this helps :-)
Also a complete weighting system must also be in place with regards to scoring. One could divide the characters into different parameters like agility, strength, charisma, etc. One could then use this for calculating the hurt scoring mechanism. I hope this helps :-)
Author
Didn't Street Fighter 2 use attack patterns? The patterns were very noticable. Like Ken doing a hurricane kick into a dragon punch a few times in a row. And him jumping back and forth trying to get you to attack him. Things like that.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
