protected void runEpisode(MetaNEATGenome genome) {
RLGlue.RL_agent_message(genome.toPhenotype().toString());
RLGlue.RL_init();
RLGlue.RL_episode(steps);
double totalReward = RLGlue.RL_return();
genome.addFitness(rewardToFitness(totalReward));
genome.incrEpisodes();
RLGlue.RL_cleanup();
}
public static void main(String[] args) {
Thread rlgluethread = new RLGlueThread();
rlgluethread.start();
EnvThread envthread = new EnvThread();
envthread.start();
AgentThread agentthread = new AgentThread();
agentthread.start();
int whichTrainingMDP = 0; // select the MDP to load
for (int it = 0; it < 20; it++) {
whichTrainingMDP = it;
consoleTrainerHelper.loadTetris(whichTrainingMDP); // whichTrainingMDP should be in [0,19]
RLGlue.RL_init();
int stepsRemaining = 50000;
int totalEpisodes = 0;
double returnThisMDP = 0.0d;
while (stepsRemaining > 0) {
RLGlue.RL_episode(stepsRemaining);
int thisStepCount = RLGlue.RL_num_steps();
stepsRemaining -= thisStepCount;
returnThisMDP += RLGlue.RL_return();
totalEpisodes++;
}
System.out.println(
"MDP "
+ it
+ " completed with "
+ totalEpisodes
+ " episodes, got "
+ returnThisMDP
+ " reward");
// clean up the environment and end the program
RLGlue.RL_cleanup();
}
System.exit(0);
}
public static void main(String[] args) {
/* Log level*/
Logger.getRootLogger().setLevel(Level.INFO);
/** ************* Setup Agent ************* */
/* create RLGlue Agent to get task specification (needed by the learners Q-function approximator) */
RlGlueAgent rlGlueAgent = new RlGlueAgent();
// create an agentLoader that will start the agent when its run method is called
AgentLoader agentLoader = new AgentLoader(rlGlueAgent);
// create thread so that the agent and environment can run asynchronously
Thread agentThread = new Thread(agentLoader);
// start the thread
agentThread.start();
String taskSpec = RLGlue.RL_init();
System.out.println("Task-Specification: " + taskSpec);
/**
* ************************************ Configure Tile-Coding approximation
* ************************************
*/
// BEGIN TILE CODING APPROXIMATION
// the number of tilings/layers
int nTilings = 5;
// the configuration. {from, to, number if discs}
double[][] config =
new double[][] {
{0, 1.0, 5},
{0, 1.0, 5},
};
// create square tilings
Network net = new Network();
net.setIsNormalized(true);
double[][] optimizationConfig = config.clone();
net.setFeatureGenerator(
new GridHashFeatureGenerator(optimizationConfig, new TileAndIndexBoundingBoxCalculator()));
net.add(TileCodingFactory.createTilings(config, nTilings));
// setup Q-Function
QFeatureFunction Q = new QFeatureFunction(net, rlGlueAgent.getTeachingboxActionSet());
// END TILE-CODING
/**
* *************************************** setup policy, learner & the TB's agent
* ***************************************
*/
// the ActionSet for the policy is read from the rlGlueAgent (RL_init must have been called
// before!)
EpsilonGreedyPolicy pi = new EpsilonGreedyPolicy(Q, rlGlueAgent.getTeachingboxActionSet(), 0.1);
System.out.println("POLICY-LEARNER ActionSet: " + rlGlueAgent.getTeachingboxActionSet());
GradientDescentSarsaLearner learner =
new GradientDescentSarsaLearner(Q, net, rlGlueAgent.getTeachingboxActionSet());
learner.setAlpha(0.5);
learner.setGamma(1.0);
learner.setLambda(0.9);
Agent tbAgent = new Agent(pi);
tbAgent.addObserver(learner);
/* SET THE TEACHINGBOX-AGENT IN THE RL-GLUE-AGENT-ADAPTER */
rlGlueAgent.setTeachingBoxAgent(tbAgent);
/**
* ******************************* Setup Experiment and Plotting *******************************
*/
RLGlueRemoteEnvironment rlEnv = new RLGlueRemoteEnvironment();
Experiment experiment = new Experiment(tbAgent, rlEnv, 100, 1000);
// 3D PLOTTING
// draw the maximum value of the QFunction
// to plot the corresponding VFunction we just have to pass in the policy
// as well as the actionSet
ValueFunctionEQ V = new ValueFunctionEQ(Q);
V.costfunction = true;
Plotter Vplotter =
new ValueFunctionSurfacePlotter(V, "[0:0.02:1.0]", "[0:0.02:1.0]", "PuddleWorld");
// use action runtime plotter, that calls the ValueFunctionPlotter every 10th episode
Vplotter = new RuntimePlotter(Vplotter, RuntimePlotter.Mode.EPISODE, 10, net);
// add the plotter as an observer to the experiment
experiment.addObserver((RuntimePlotter) Vplotter);
// RUN THE EXPERIMENT
experiment.run();
// cleanup rl-glue at the end
RLGlue.RL_cleanup();
System.exit(1);
}
public void runExperiment() {
System.out.println("\n\nExperiment starting up!");
String taskSpec = RLGlue.RL_init();
System.out.println("RL_init called, the environment sent task spec: " + taskSpec);
System.out.println("\n\n----------Sending some sample messages----------");
/* Talk to the agent and environment a bit... */
String responseMessage = RLGlue.RL_agent_message("what is your name?");
System.out.println("Agent responded to \"what is your name?\" with: " + responseMessage);
responseMessage = RLGlue.RL_agent_message("If at first you don't succeed; call it version 1.0");
System.out.println(
"Agent responded to \"If at first you don't succeed; call it version 1.0 \" with: "
+ responseMessage
+ "\n");
responseMessage = RLGlue.RL_env_message("what is your name?");
System.out.println("Environment responded to \"what is your name?\" with: " + responseMessage);
responseMessage = RLGlue.RL_env_message("If at first you don't succeed; call it version 1.0");
System.out.println(
"Environment responded to \"If at first you don't succeed; call it version 1.0 \" with: "
+ responseMessage);
System.out.println("\n\n----------Running a few episodes----------");
for (int i = 0; i < 1; i++) runEpisode(20000);
runEpisode(1);
/* Remember that stepLimit of 0 means there is no limit at all! */
// runEpisode(0);
RLGlue.RL_cleanup();
System.out.println("\n\n----------Stepping through an episode----------");
// taskSpec = RLGlue.RL_init();
// for(int i = 0; i < 2; i++) {
// /*We could also start over and do another experiment */
// /*We could run one step at a time instead of one episode at a time */
// /*Start the episode */
// Observation_action startResponse = RLGlue.RL_start();
// int firstObservation = startResponse.o.intArray[0];
// int firstAction = startResponse.a.intArray[0];
// System.out.println("First observation and action were: " +
// firstObservation + " and: " + firstAction);
//
// /*Run one step */
// Reward_observation_action_terminal stepResponse = RLGlue.RL_step();
//
// /*Run until the episode ends*/
// while (stepResponse.terminal != 1) {
// stepResponse = RLGlue.RL_step();
// if (stepResponse.terminal != 1) {
// /*Could optionally print state,action pairs */
// System.out.println(
// ("(state,action)=(" +
// stepResponse.o.intArray[0] +
// "," +
// stepResponse.a.intArray[0] + ")"));
// }
// }
//
// System.out.println("\n\n----------Summary----------");
//
// int totalSteps = RLGlue.RL_num_steps();
// double totalReward = RLGlue.RL_return();
// System.out.println("It ran for " + totalSteps +
// " steps, total reward was: " + totalReward);
// }
// RLGlue.RL_cleanup();
}