RLLearner::RLLearner(ALEInterface& ale, Parameters *param){
frame = 0;
randomActionTaken = 0;
gamma = param->getGamma();
epsilon = param->getEpsilon();
toUseOnlyRewardSign = param->getUseRewardSign();
numStepsPerAction = param->getNumStepsPerAction();
toBeOptimistic = param->getOptimisticInitialization();
episodeLength = param->getEpisodeLength();
numEpisodesEval = param->getNumEpisodesEval();
numEpisodesLearn = param->getNumEpisodesLearn();
//Get the number of effective actions:
if(param->isMinimalAction()){
actions = ale.getMinimalActionSet();
}
else{
actions = ale.getLegalActionSet();
}
numActions = actions.size();
}
int main(int argc, char** argv) {
ALEInterface ale;
// Get & Set the desired settings
ale.setInt("random_seed", 123);
//The default is now 0 because we don't want stochasity
ale.setFloat("repeat_action_probability", 0);
#ifdef __USE_SDL
ale.setBool("display_screen", false);
ale.setBool("sound", false);
#endif
/// Uncomment to Record
// std::string recordPath = "record";
// std::cout << std::endl;
// // Set record flags
// ale.setString("record_screen_dir", recordPath.c_str());
// ale.setString("record_sound_filename", (recordPath + "/sound.wav").c_str());
// // We set fragsize to 64 to ensure proper sound sync
// ale.setInt("fragsize", 64);
// // Not completely portable, but will work in most cases
// std::string cmd = "mkdir ";
// cmd += recordPath;
// system(cmd.c_str());
// Load the ROM file. (Also resets the system for new settings to
// take effect.)
ale.loadROM("gravitar.bin");
// Get the vector of minimal actions
const ActionVect minimal_actions = ale.getMinimalActionSet();
// Erase actions that move, but don't fire
//minimal_actions.erase(minimal_actions.begin() + 2, minimal_actions.begin() + 10);
// Store all rewards earned in all episodes
float allRewards = 0;
double allTimes = 0;
Timer timer;
// Play 10 episodes
int episodes = 200;
int number = 0;
int count = 0;
int lastLives = ale.lives();
bool reset = false;
Decision decision = Decision(ale.getMinimalActionSet(), ale.getScreen());
for (int episode=0; episode<episodes; episode++) {
float totalReward = 0;
double episodeTime = 0;
timer.start();
while (!ale.game_over()) {
if (ale.lives() < lastLives){
lastLives = ale.lives();
number = 0;
count = 0;
reset = true;
//cout << " DIE " << endl;
} else{
reset = false;
}
// Apply the action and get the resulting reward
float reward = ale.act(decision.getDecision(ale.getScreen(), ale.lives(), reset));
//decision.print();
totalReward += reward;
}
timer.stop();
episodeTime = timer();
timer.reset();
count = 0;
number = 0;
allRewards += totalReward;
allTimes += episodeTime;
cout << "Episode " << episode << " ended with score: " << totalReward << " with time: "<< episodeTime <<endl;
ale.reset_game();
}
// Display average reward per game
cout << "Average Reward: " << (allRewards / episodes) << " Average Time: " << (allTimes/episodes) << endl;
return 0;
}