OffPolicyAgentDirect.java

Go to the documentation of this file.

package rlpark.plugin.rltoys.agents.offpolicy;

import rlpark.plugin.rltoys.agents.rl.ControlAgent;
import rlpark.plugin.rltoys.algorithms.control.OffPolicyLearner;
import rlpark.plugin.rltoys.envio.actions.Action;
import rlpark.plugin.rltoys.envio.policy.Policies;
import rlpark.plugin.rltoys.envio.policy.Policy;
import rlpark.plugin.rltoys.envio.rl.RLAgent;
import rlpark.plugin.rltoys.envio.rl.TRStep;
import rlpark.plugin.rltoys.math.vector.RealVector;
import rlpark.plugin.rltoys.math.vector.implementations.PVector;
import rlpark.plugin.rltoys.math.vector.implementations.VectorNull;
import zephyr.plugin.core.api.monitoring.annotations.Monitor;

@Monitor
public class OffPolicyAgentDirect implements OffPolicyAgentEvaluable {
  private static final long serialVersionUID = -8255597969677460009L;
  private RealVector x_t;
  private final OffPolicyLearner learner;
  private final Policy behaviour;

  public OffPolicyAgentDirect(Policy behaviour, OffPolicyLearner learner) {
    this.learner = learner;
    this.behaviour = behaviour;
  }

  @Override
  public Action getAtp1(TRStep step) {
    if (step.isEpisodeStarting())
      x_t = null;
    RealVector x_tp1 = step.o_tp1 != null ? new PVector(step.o_tp1) : new VectorNull(x_t.getDimension());
    Action a_tp1 = step.o_tp1 != null ? Policies.decide(behaviour, x_tp1) : null;
    learner.learn(x_t, step.a_t, x_tp1, a_tp1, step.r_tp1);
    x_t = x_tp1;
    return a_tp1;
  }

  public OffPolicyLearner learner() {
    return learner;
  }

  public Policy behaviour() {
    return behaviour;
  }

  @Override
  public RLAgent createEvaluatedAgent() {
    return new ControlAgent(learner);
  }
}