First commit

Signed-off-by: christian peña <christian@visualprogress.ai>

README.md

@@ -1,44 +1,2 @@
-This project was bootstrapped with [Create React App](https://github.com/facebook/create-react-app).
 
-## Available Scripts
-
-In the project directory, you can run:
-
-### `npm start`
-
-Runs the app in the development mode.<br>
-Open [http://localhost:3000](http://localhost:3000) to view it in the browser.
-
-The page will reload if you make edits.<br>
-You will also see any lint errors in the console.
-
-### `npm test`
-
-Launches the test runner in the interactive watch mode.<br>
-See the section about [running tests](https://facebook.github.io/create-react-app/docs/running-tests) for more information.
-
-### `npm run build`
-
-Builds the app for production to the `build` folder.<br>
-It correctly bundles React in production mode and optimizes the build for the best performance.
-
-The build is minified and the filenames include the hashes.<br>
-Your app is ready to be deployed!
-
-See the section about [deployment](https://facebook.github.io/create-react-app/docs/deployment) for more information.
-
-### `npm run eject`
-
-**Note: this is a one-way operation. Once you `eject`, you can’t go back!**
-
-If you aren’t satisfied with the build tool and configuration choices, you can `eject` at any time. This command will remove the single build dependency from your project.
-
-Instead, it will copy all the configuration files and the transitive dependencies (Webpack, Babel, ESLint, etc) right into your project so you have full control over them. All of the commands except `eject` will still work, but they will point to the copied scripts so you can tweak them. At this point you’re on your own.
-
-You don’t have to ever use `eject`. The curated feature set is suitable for small and middle deployments, and you shouldn’t feel obligated to use this feature. However we understand that this tool wouldn’t be useful if you couldn’t customize it when you are ready for it.
-
-## Learn More
-
-You can learn more in the [Create React App documentation](https://facebook.github.io/create-react-app/docs/getting-started).
-
-To learn React, check out the [React documentation](https://reactjs.org/).
+# Sneik

libs/Brain.js

@@ -0,0 +1,294 @@
+const Settings = require('./Settings')
+const tf = require('@tensorflow/tfjs')
+
+const NUM_ACTIONS = 4
+
+class Brain {
+  constructor (memory) {
+    this.generateActivations = true
+    this.memory = memory
+    this.inputShape = [Settings.BOARD_WIDTH, Settings.BOARD_HEIGHT, 1]
+    this.epsilon = Settings.EPSILON
+    this.Q_table = new Array(NUM_ACTIONS)
+    this.activationMaps = {}
+    this.states = []
+  }
+
+  load () {
+    return tf.io.listModels()
+      .then((list) => {
+        let modelFilename = Settings.SAVE_MODEL_DESTINATION + Settings.MODEL_NAME
+        console.log('Available models', list, modelFilename)
+        if (Object.keys(list).indexOf(modelFilename) !== -1) {
+          return this._loadModel()
+        } else {
+          return this._buildModel()
+        }
+      })
+      .then(() => {
+        if (Settings.TRAINING_INTERVAL > 0) {
+          this.trainInterval = setInterval(() => {
+            this._trainModel()
+          }, Settings.TRAINING_INTERVAL)
+        }
+      })
+  }
+
+  stop () {
+    if (this.trainInterval) {
+      clearInterval(this.trainInterval)
+    }
+  }
+
+  thinkAction (state) {
+    // console.log(state)
+    let prediction = { action: null, q: null }
+
+    if (!state.willCrash) {
+      prediction = this._runDeepQLearning(state)
+    }
+    return prediction
+  }
+
+  getStats () {
+    return {
+      Q_table: this.Q_table,
+      activationMaps: this.activationMaps
+    }
+  }
+
+  // Private methods
+  async _loadModel () {
+    console.log('Loading model')
+    let modelName = Settings.MODEL_NAME
+    this.model = await tf.loadModel(Settings.SAVE_MODEL_DESTINATION + modelName)
+    this.model.compile({
+      loss: 'meanSquaredError',
+      optimizer: tf.train.sgd(Settings.LEARNING_RATE),
+      metrics: ['accuracy']
+    })
+
+    this.model.summary()
+  }
+
+  _buildModel () {
+    console.log('Building model')
+    this.convLayers = []
+    this.model = tf.sequential()
+
+    this.model.add(tf.layers.conv2d({
+      inputShape: this.inputShape,
+      filters: 16,
+      kernelSize: 5,
+      strides: 2,
+      padding: 'same',
+      activation: 'relu',
+      kernelInitializer: 'VarianceScaling' }))
+
+    this.model.add(tf.layers.conv2d({
+      filters: 32,
+      kernelSize: 3,
+      strides: 2,
+      padding: 'same',
+      activation: 'relu',
+      kernelInitializer: 'VarianceScaling' }))
+
+    this.model.add(tf.layers.flatten())
+
+    this.model.add(tf.layers.dense({
+      kernelInitializer: 'VarianceScaling',
+      units: 256
+    }))
+
+    this.model.add(tf.layers.dense({
+      units: NUM_ACTIONS,
+      kernelInitializer: 'VarianceScaling'
+    }))
+
+    this.model.compile({
+      loss: 'meanSquaredError',
+      optimizer: tf.train.sgd(Settings.LEARNING_RATE),
+      metrics: ['accuracy']
+    })
+
+    this.model.summary()
+  }
+
+  _saveModel () {
+    console.log('Saving model')
+    let modelName = Settings.MODEL_NAME
+    this.model.save(Settings.SAVE_MODEL_DESTINATION + modelName) // ('localstorage://' + modelName)
+      .then(() => {
+        console.log('Saved model!')
+      })
+      .catch(console.error)
+  }
+
+  async _runDeepQLearning (state) {
+    let prediction = { action: null, q: null }
+
+    if (Math.random() <= this.epsilon) {
+      // Random explore
+      prediction.action = Math.floor(Math.random() * NUM_ACTIONS)
+    } else {
+      // Brain explore
+      prediction = await this._makePrediction(state)
+    }
+
+    // Decay of random explore
+    if (this.epsilon > Settings.FINAL_EPSILON) {
+      this.epsilon = this.epsilon * Settings.EPSILON_DECAY
+    } else {
+      this.epsilon = Settings.FINAL_EPSILON
+    }
+
+    return prediction
+  }
+
+  async _makePrediction (state) {
+    // Creating input vector
+    let input = this._getTFInput(state.currentBoard)
+    input = input.expandDims(0)
+    // console.log(input.shape)
+
+    if (this.generateActivations) {
+      // Run model layer by layer to store activations
+      this.Q_table = await this._predictWithActivations(input)
+    } else {
+      // Run model
+      this.Q_table = await this.model.predict(input).data()
+    }
+
+    // Get action: index of max value in Q table
+    let action = (await tf.argMax(this.Q_table).data())[0]
+    // console.log('Q-table: ', this.Q_table, 'action: ', action)
+
+    // cleanup memory
+    input.dispose()
+    return { action, q: this.Q_table }
+  }
+
+  async _trainModel () {
+    console.log('Training')
+
+    // get random batch of states
+    let states = await this.memory.batch(Settings.BATCH_SIZE)
+    if (states.length === 0) {
+      console.warn('No enough data!')
+      return
+    }
+    // console.log(states)
+
+    // Prepare inputs and targets
+    let inputs = []
+    let targets = []
+    for (let i in states) {
+      let state = states[i]
+      // Get input from bpard
+      inputs.push(this._getTFInput(state.currentBoard))
+
+      // Calc reward
+      let reward = this._calcReward(state)
+
+      // Calc total reward estimated
+      let totalReward = 0
+      if (state.willCrash) {
+        totalReward = reward
+      } else {
+        let nextInput = this._getTFInput(state.nextBoard).expandDims(0)
+        let nextQtable = await this.model.predict(nextInput).data()
+        let maxNextReward = await tf.max(nextQtable).data()
+        totalReward = reward + Settings.QLEARN_GAMMA * maxNextReward
+
+        nextInput.dispose()
+      }
+
+      let target = Array.prototype.slice.call(state.q)
+      target[state.lastAction] = totalReward
+      // console.log(state.q, target, state.lastAction, totalReward)
+      targets.push(tf.tensor1d(target))
+    }
+    inputs = tf.stack(inputs)
+    targets = tf.stack(targets)
+
+    // Train model
+    this.model.fit(inputs, targets, {
+      batchSize: states.length,
+      epochs: 1,
+      callbacks: {
+        onBatchEnd: () => console.log('Training ended!')
+      }
+    })
+      .then(() => {
+        // Cleanup memory
+        inputs.dispose()
+        targets.dispose()
+
+        // Save model
+        if (Settings.SAVE_MODEL) {
+          this._saveModel()
+        }
+      })
+      .catch(console.error)
+  }
+
+  async _predictWithActivations (input) {
+    let inputs = []
+    inputs.push(input)
+    for (var i = 0; i < this.model.layers.length; i++) {
+      inputs.push(this.model.layers[i].apply(inputs[i]))
+
+      // Get activations in conv2D layers
+      if (this.model.layers[i].name.indexOf('conv2d') !== -1) {
+        const shape = inputs[i + 1].shape
+        const convList = tf.unstack(inputs[i + 1].reshape([shape[1], shape[2], shape[3]]), 2)
+        this.activationMaps[this.model.layers[i].name] = []
+        for (let j in convList) {
+          this.activationMaps[this.model.layers[i].name].push({ map: convList[j].dataSync(), shape: convList[j].shape })
+          convList[j].dispose()
+        }
+      }
+    }
+    let q = await inputs[inputs.length - 1].data()
+
+    for (let i = inputs.length - 1; i >= 0; i--) {
+      inputs[i].dispose()
+    }
+    return q
+  }
+
+  _calcReward (state) {
+    let reward = Settings.QLEARN_REWARD_ALIVE
+    if (Settings.QLEARN_NEARNESS_REWARD) {
+      if (state.snakePos && state.foodPos) {
+        let newDistance = this._getDistanceToFood(state.snakePos, state.foodPos)
+        if (this.oldDistance) {
+          if (newDistance < this.oldDistance) {
+            reward = Settings.QLEARN_REWARD_CLOSE
+          } else {
+            reward = Settings.QLEARN_REWARD_FAR
+          }
+        }
+        this.oldDistance = newDistance
+      }
+    }
+    if (state.willCrash) {
+      reward = Settings.QLEARN_REWARD_DIE
+    } else if (state.willEat) {
+      reward = Settings.QLEARN_REWARD_EAT
+    }
+    return reward
+  }
+
+  _getDistanceToFood (snake, food) {
+    let x = Math.pow((snake[0] - food[0]), 2)
+    let y = Math.pow((snake[1] - food[1]), 2)
+    return Math.sqrt(x + y)
+  }
+
+  _getTFInput (board) {
+    return tf.tensor2d(board).div(255.0).expandDims(2)
+  }
+}
+
+module.exports = Brain