Method playStep

snake-dqn/agent.js:81–121

* Play one step of the game. * * @returns {number | null} If this step leads to the end of the game, * the total reward from the game as a plain number. Else, `null`.

()

   *   the total reward from the game as a plain number. Else, `null`.
   */
  playStep() {
    this.epsilon = this.frameCount >= this.epsilonDecayFrames ?
        this.epsilonFinal :
        this.epsilonInit + this.epsilonIncrement_  * this.frameCount;
    this.frameCount++;

    // The epsilon-greedy algorithm.
    let action;
    const state = this.game.getState();
    if (Math.random() < this.epsilon) {
      // Pick an action at random.
      action = getRandomAction();
    } else {
      // Greedily pick an action based on online DQN output.
      tf.tidy(() => {
        const stateTensor =
            getStateTensor(state, this.game.height, this.game.width)
        action = ALL_ACTIONS[
            this.onlineNetwork.predict(stateTensor).argMax(-1).dataSync()[0]];
      });
    }

    const {state: nextState, reward, done, fruitEaten} = this.game.step(action);

    this.replayMemory.append([state, action, reward, done, nextState]);

    this.cumulativeReward_ += reward;
    if (fruitEaten) {
      this.fruitsEaten_++;
    }
    const output = {
      action,
      cumulativeReward: this.cumulativeReward_,
      done,
      fruitsEaten: this.fruitsEaten_
    };
    if (done) {
      this.reset();
    }
    return output;
  }

  /**
   * Perform training on a randomly sampled batch from the replay buffer.