int main(int argc, char** argv) {
if (argc < 2) {
std::cerr << "Usage: " << argv[0] << " rom_file" << std::endl;
return 1;
}
ALEInterface ale;
// Get & Set the desired settings
ale.setInt("random_seed", 123);
//The default is already 0.25, this is just an example
ale.setFloat("repeat_action_probability", 0.25);
#ifdef __USE_SDL
ale.setBool("display_screen", true);
ale.setBool("sound", true);
#endif
// Load the ROM file. (Also resets the system for new settings to
// take effect.)
ale.loadROM(argv[1]);
// Get the vector of legal actions
ActionVect legal_actions = ale.getLegalActionSet();
// Play 10 episodes
for (int episode=0; episode<1; episode++) {
float totalReward = 0;
while (!ale.game_over()) {
/*
const ALEScreen screen = ale.getScreen();
float row_sum = 0, column_sum = 0, tot = 0.0;
for (int i = 0; i < screen.height(); ++i)
{
for (int j = 0; j < screen.width(); ++j)
{
pixel_t tmp = screen.get(i, j);
if(tmp >= 41 && tmp <= 43) {
row_sum += i;
column_sum += j;
tot++;
}
}
}
cout << (row_sum/tot) << "," << (column_sum/tot) << endl;
*/
Action a = legal_actions[rand() % legal_actions.size()];
// Apply the action and get the resulting reward
float reward = ale.act(a);
totalReward += reward;
}
cout << "Episode " << episode << " ended with score: " << totalReward << endl;
ale.reset_game();
}
return 0;
}
int main(int argc, char** argv) {
if (argc < 2) {
std::cout << "Usage: " << argv[0] << " rom_file" << std::endl;
return 1;
}
ALEInterface ale;
// Get & Set the desired settings
ale.setInt("random_seed", 123);
// We enable both screen and sound, which we will need for recording.
ale.setBool("display_screen", true);
// You may leave sound disabled (by setting this flag to false) if so desired.
ale.setBool("sound", true);
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(argv[1]);
// Get the vector of legal actions
ActionVect legal_actions = ale.getLegalActionSet();
// Play a single episode, which we record.
while (!ale.game_over()) {
Action a = legal_actions[rand() % legal_actions.size()];
// Apply the action (discard the resulting reward)
ale.act(a);
}
std::cout << std::endl;
std::cout << "Recording complete. To create a video, you may want to run \n"
" doc/scripts/videoRecordingExampleJoinXXX.sh. See manual for details.." << std::endl;
return 0;
}
//Run the Arcade Learning Environment using the DQN agent.
void run_ale(int argc, char** argv)
{
//Create Arcade Learning Environment
ALEInterface* ale = new ALEInterface(false);
//Load the Atari Rom we are going to play
ale->loadROM(argv[1]);
//Get the set of possible actions from ALE ROM
ActionVect action_set = {PLAYER_A_LEFTFIRE,PLAYER_A_FIRE,PLAYER_A_RIGHTFIRE}; //ale->getMinimalActionSet();
//Create action descriptor
ActionDescriptor descriptor({action_set.size()},{});
//Create Learning System
DACN system(descriptor,0.9,4,1000000,100,32);
//Set exploration rate
system.exploration_rate(0.9);
cudaProfilerStart(); nvtxRangePushA("2 Atari Games");
for(int episode=0; episode<2; episode++)
{
string tmp = "episode: " + to_string(episode);
nvtxRangePushA(tmp.c_str());
//Restart the game
ale->reset_game();
//Game Loop
while(!ale->game_over())
{
nvtxRangePushA("step");
//Convert screen to input
ALEScreen screen = ale->getScreen();
gray8_image_t img = to_image(screen);
vector<unsigned char> input = to_input(img);
float raw_action = system.forward( input )[0];
//cast action
int action = static_cast<int>(raw_action);
//Execute the action and get the reward
float reward = ale->act(action_set[action]);
//Normalize the reward
float normalized_reward = max(min(1.0f,reward),-1.0f);
//Backward the result
system.backward(normalized_reward,ale->game_over());
nvtxRangePop();
}
nvtxRangePop();
}
cudaProfilerStop(); nvtxRangePop();
}
int main(int argc, char** argv) {
if (argc < 2) {
std::cerr << "Usage: " << argv[0] << " rom_file" << std::endl;
return 1;
}
ALEInterface ale;
// Get & Set the desired settings
ale.setInt("random_seed", 123);
//The default is already 0.25, this is just an example
ale.setFloat("repeat_action_probability", 0.25);
#ifdef __USE_SDL
ale.setBool("display_screen", true);
ale.setBool("sound", true);
#endif
// Load the ROM file. (Also resets the system for new settings to
// take effect.)
ale.loadROM(argv[1]);
// Get the vector of legal actions
ActionVect legal_actions = ale.getLegalActionSet();
// Play 10 episodes
for (int episode=0; episode<10; episode++) {
float totalReward = 0;
while (!ale.game_over()) {
Action a = legal_actions[rand() % legal_actions.size()];
// Apply the action and get the resulting reward
float reward = ale.act(a);
totalReward += reward;
if (reward != 0){
cout<<"totalReward: "<<totalReward<<endl;
}
}
cout << "Episode " << episode << " ended with score: " << totalReward << endl;
ale.reset_game();
}
return 0;
}
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;
}