Consider that there are essentially infinite combinations of keypresses in a game like SMB, per frame
well, for every frame, there's 2^8 input possibilities, or 256. If you take a video that's 10 minutes long, that's 60fps * 600 sec = 36000 frames, giving you a total number of 9,216,000 different movie possibilities. If you can cull the amount of control input, say, no up+down; or if you give it instructions to watch for certain RAM values- no death, don't pick up a certain item; you could reduce that number to something more manageable.
There's actually a factorial amount of them (with respect to the number of pressable keys), which is far fewer than infinite. ;)
This is an interesting idea. Many turn-based strategy games involve a lot of luck, so the warbot could also be programmed to take the enemy's turn by means of manipulation. For example, moving one of your units might make the AI do something totally different if done one frame later, causing the entire branch to be "incorrect".
I think your best bet would be to actually dissamble the rom and try to simulate the game. This would make programming much more easy and of course faster.
What sorts of framerates do you guys get? [in Snes9x]
I'm not sure if you understood what I meant, but the idea would be to recreate the enemy AI in your own program, so you wouldn't need to use the emulator at all. This custom AI would be faster for bruteforcing purposes since it doesn't calculate useless things like graphics, sound etc.
Not really sure if this is accurate, but SNES9x seems to indicate that I get about 59 FPS for NTSC games and around 49 for PAL. (Running a P4@3GHz w/2GB RAM & Radeon 9200SE@128MB)
Maximus: you should try turning off the 'Automatic Frame Skip' option, that removes the cap from frame rate.
private void testallactions(int round){ // round=1 to start at round 1
for(int i=0;i<5;i++){ // loop through all possible actions
state[round]=state[round-1]; // load last round's state
state[round].action(i); // execute action #i
command[round]=i; // log the executed action
if(state[round].result==0 && round<maxround){ // no one died and max roundnumber is not reached
testallactions(round+1); // test all possible actions in the next round
}else{
for(int j=1;j<=round;j++){
System.out.print(command[j] + " "); // print all executed actions
}
if(state[round].result==1){ // maxim died
System.out.println("Maxim died.");
}else if(state[round].result==2){ // catfish died, no jewel
System.out.println("Catfish died - no jewel.");
}else if(state[round].result==3){ // catfish died, jewel dropped
System.out.println("Catfish died - jewel dropped.");
}
} // try next action
}
}state[round]=state[round-1]; // load last round's state
The first thing I did was building a simulator, which turned out to be able to simulate ~500000 battle rounds per second.
public void action(int n){
rng.wasteRN(6); // agility check (but Maxim always wins)
int def = 2; // defence modifier
int rngwaste = 0; // delayed RNG calls
if(n==0){ // defend
def = 1;
}else if(n==1){ // attack
int damage = (rng.callRN()/4*maxim[3]/256+maxim[3]-10/2)/2;
rng.wasteRN(20);
cfhealth = cfhealth - damage;
}else if(n==2){ // dive (2x damage)
int damage = (rng.callRN()/4*maxim[3]/256+maxim[3]-10/2);
rng.wasteRN(20);
cfhealth = cfhealth - damage;
}else if(n==3){ // boomerang ("ATP" = 100)
while(rng.callRN()>42){}
int damage = (rng.callRN()/8*100/256+100-10/2)/2;
rng.wasteRN(20);
rngwaste = 30;
cfhealth = cfhealth - damage;
}else if(n==4){ // potion ("ATP" = 30)
int heal = rng.callRN()*11/256*30/256+30;
rng.wasteRN(20);
maxim[0] = maxim[0] + heal;
if(maxim[0]>maxim[1]){maxim[0] = maxim[1];} // check if HP does not exceed max HP
}
if(cfhealth<1){ // if the catfish died:
if(rng.callRN()<64>63 && a<128){ // secundary item drop check
result = 3; // item drop checks succeeded
}else{
result = 2; // item drop checks failed
}
}
}else{ // if the catfish lives:
if(rngwaste!=0){ // call the RNG during spell animation (after item drop checks)
rng.wasteRN(rngwaste);
rngwaste=0;
}
AI(def); // Catfish' turn
if(maxim[0]<1){
result = 1;
}
}
}