mountain_car.hpp

Go to the documentation of this file.

#ifndef MLPACK_METHODS_RL_ENVIRONMENT_MOUNTAIN_CAR_HPP
#define MLPACK_METHODS_RL_ENVIRONMENT_MOUNTAIN_CAR_HPP

#include <mlpack/prereqs.hpp>
#include <mlpack/core/math/clamp.hpp>

namespace mlpack {
namespace rl {

class MountainCar
{
 public:
  class State
  {
   public:
    State(): data(dimension, arma::fill::zeros)
    { /* Nothing to do here. */ }

    State(const arma::colvec& data): data(data)
    { /* Nothing to do here. */ }

    arma::colvec& Data() { return data; }

    double Velocity() const { return data[0]; }
    double& Velocity() { return data[0]; }

    double Position() const { return data[1]; }
    double& Position() { return data[1]; }

    const arma::colvec& Encode() const { return data; }

    static constexpr size_t dimension = 2;

   private:
    arma::colvec data;
  };

  class Action
  {
   public:
    enum actions
    {
      backward,
      stop,
      forward
    };
    // To store the action.
    Action::actions action;

    // Track the size of the action space.
    static const size_t size = 3;
  };

  MountainCar(const size_t maxSteps = 200,
              const double positionMin = -1.2,
              const double positionMax = 0.6,
              const double positionGoal = 0.5,
              const double velocityMin = -0.07,
              const double velocityMax = 0.07,
              const double doneReward = 0) :
      maxSteps(maxSteps),
      positionMin(positionMin),
      positionMax(positionMax),
      positionGoal(positionGoal),
      velocityMin(velocityMin),
      velocityMax(velocityMax),
      doneReward(doneReward),
      stepsPerformed(0)
  { /* Nothing to do here */ }

  double Sample(const State& state,
                const Action& action,
                State& nextState)
  {
    // Update the number of steps performed.
    stepsPerformed++;

    // Calculate acceleration.
    int direction = action.action - 1;
    nextState.Velocity() = state.Velocity() + 0.001 * direction - 0.0025 *
        std::cos(3 * state.Position());
    nextState.Velocity() = math::ClampRange(nextState.Velocity(),
        velocityMin, velocityMax);

    // Update states.
    nextState.Position() = state.Position() + nextState.Velocity();
    nextState.Position() = math::ClampRange(nextState.Position(),
        positionMin, positionMax);

    if (nextState.Position() == positionMin && nextState.Velocity() < 0)
      nextState.Velocity() = 0.0;

    // Check if the episode has terminated.
    bool done = IsTerminal(nextState);

    // Do not reward the agent if time ran out.
    if (done && maxSteps != 0 && stepsPerformed >= maxSteps)
      return 0;
    else if (done)
      return doneReward;

    return -1;
  }

  double Sample(const State& state, const Action& action)
  {
    State nextState;
    return Sample(state, action, nextState);
  }

  State InitialSample()
  {
    State state;
    stepsPerformed = 0;
    state.Velocity() = 0.0;
    state.Position() = arma::as_scalar(arma::randu(1)) * 0.2 - 0.6;
    return state;
  }

  bool IsTerminal(const State& state) const
  {
    if (maxSteps != 0 && stepsPerformed >= maxSteps)
    {
      Log::Info << "Episode terminated due to the maximum number of steps"
          "being taken.";
      return true;
    }
    else if (state.Position() >= positionGoal)
    {
      Log::Info << "Episode terminated due to agent succeeding.";
      return true;
    }
    return false;
  }

  size_t StepsPerformed() const { return stepsPerformed; }

  size_t MaxSteps() const { return maxSteps; }
  size_t& MaxSteps() { return maxSteps; }

 private:
  size_t maxSteps;

  double positionMin;

  double positionMax;

  double positionGoal;

  double velocityMin;

  double velocityMax;

  double doneReward;

  size_t stepsPerformed;
};

} // namespace rl
} // namespace mlpack

#endif