Learner.h

#pragma once

#include <vector>

namespace net {
class NeuralNet;
}

namespace rl {

typedef std::vector<double> Action;
typedef std::vector<double> State;

/** A reinforcement learning system. */
class Learner {
public:
  /**
   * \brief Gets the action that the network deems most beneficial for the currentState
   *
   * \param currentState the state for which to choose the action
   */
  virtual Action chooseBestAction(State currentState) = 0;

  /**
   * \brief Gets an action using the Boltzman softmax probability distribution
   *
   * A non-random search heuristic is used such that the neural network explores actions despite their reward value.
   * The lower the exploration constant, the more likely it is to pick the best action for the current state.
   *
   * \param currentState the state for which to choose the action
   * \param explorationConstant the Boltzmann temperature constant, determining "exploration"
   */
  virtual Action chooseBoltzmanAction(State currentState, double explorationConstant) = 0;

  /**
   * \brief Apply reinforcement to the last action
   *
   * Given the immediate reward from the last action taken and the new state,
   * this function updates the correct value for the longterm reward of the lastAction and trains the network in charge of the lastAction to output the correct reward value
   *
   * \param reward the reward given for the last action taken
   * \param newState the new state
   */
  virtual void applyReinforcementToLastAction(double reward, State newState) = 0;

  /** Randomizes the leaner.  */
  virtual void reset() = 0;
};
};