/**
* The copyright in this software is being made available under the BSD License,
* included below. This software may be subject to other third party and contributor
* rights, including patent rights, and no such rights are granted under this license.
*
* Copyright (c) 2013, Dash Industry Forum.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification,
* are permitted provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation and/or
* other materials provided with the distribution.
* * Neither the name of Dash Industry Forum nor the names of its
* contributors may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS AS IS AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
* INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/**
* Authors:
* Abdelhak Bentaleb | National University of Singapore | bentaleb@comp.nus.edu.sg
* Mehmet N. Akcay | Ozyegin University | necmettin.akcay@ozu.edu.tr
* May Lim | National University of Singapore | maylim@comp.nus.edu.sg
*/
import FactoryMaker from '../../../../core/FactoryMaker';
import Debug from '../../../../core/Debug';
const WEIGHT_SELECTION_MODES = {
MANUAL: 'manual_weight_selection',
RANDOM: 'random_weight_selection',
DYNAMIC: 'dynamic_weight_selection'
};
function LearningAbrController() {
const context = this.context;
let instance,
logger,
somBitrateNeurons,
bitrateNormalizationFactor,
latencyNormalizationFactor,
minBitrate,
weights,
sortedCenters,
weightSelectionMode;
/**
* Setup the class
*/
function _setup() {
logger = Debug(context).getInstance().getLogger(instance);
_resetInitialSettings();
}
/**
* Reset all values
*/
function reset() {
_resetInitialSettings();
}
/**
* Reset to initial settings
* @private
*/
function _resetInitialSettings() {
somBitrateNeurons = null;
bitrateNormalizationFactor = 1;
latencyNormalizationFactor = 100;
minBitrate = 0;
weights = null;
sortedCenters = null;
weightSelectionMode = WEIGHT_SELECTION_MODES.DYNAMIC;
}
/**
* Returns the maximum throughput
* @return {number}
* @private
*/
function _getMaxThroughput() {
let maxThroughput = 0;
if (somBitrateNeurons) {
for (let i = 0; i < somBitrateNeurons.length; i++) {
let neuron = somBitrateNeurons[i];
if (neuron.state.throughput > maxThroughput) {
maxThroughput = neuron.state.throughput;
}
}
}
return maxThroughput;
}
/**
*
* @param {array} w
* @return {number}
* @private
*/
function _getMagnitude(w) {
const magnitude = w.map((x) => (Math.pow(x, 2))).reduce((sum, now) => sum + now);
return Math.sqrt(magnitude);
}
/**
*
* @param {array} a
* @param {array} b
* @param {array} w
* @return {number}
* @private
*/
function _getDistance(a, b, w) {
let sum = a
.map((x, i) => (w[i] * (Math.pow(x - b[i], 2)))) // square the difference*w
.reduce((sum, now) => sum + now); // sum
let sign = (sum < 0) ? -1 : 1;
return sign * Math.sqrt(Math.abs(sum));
}
/**
*
* @param {object} a
* @param {object} b
* @return {number}
* @private
*/
function _getNeuronDistance(a, b) {
let aState = [a.state.throughput, a.state.latency, a.state.rebuffer, a.state.switch];
let bState = [b.state.throughput, b.state.latency, b.state.rebuffer, b.state.switch];
return _getDistance(aState, bState, [1, 1, 1, 1]);
}
/**
*
* @param {object} winnerNeuron
* @param {array} somElements
* @param {array} x
* @private
*/
function _updateNeurons(winnerNeuron, somElements, x) {
for (let i = 0; i < somElements.length; i++) {
let somNeuron = somElements[i];
let sigma = 0.1;
const neuronDistance = _getNeuronDistance(somNeuron, winnerNeuron);
let neighbourHood = Math.exp(-1 * Math.pow(neuronDistance, 2) / (2 * Math.pow(sigma, 2)));
_updateNeuronState(somNeuron, x, neighbourHood);
}
}
/**
*
* @param {object} neuron
* @param {array} x
* @param {object} neighbourHood
* @private
*/
function _updateNeuronState(neuron, x, neighbourHood) {
let state = neuron.state;
let w = [0.01, 0.01, 0.01, 0.01]; // learning rate
state.throughput = state.throughput + (x[0] - state.throughput) * w[0] * neighbourHood;
state.latency = state.latency + (x[1] - state.latency) * w[1] * neighbourHood;
state.rebuffer = state.rebuffer + (x[2] - state.rebuffer) * w[2] * neighbourHood;
state.switch = state.switch + (x[3] - state.switch) * w[3] * neighbourHood;
}
/**
*
* @param {object} currentNeuron
* @param {number} currentThroughput
* @return {object}
* @private
*/
function _getDownShiftNeuron(currentNeuron, currentThroughput) {
let maxSuitableBitrate = 0;
let result = currentNeuron;
if (somBitrateNeurons) {
for (let i = 0; i < somBitrateNeurons.length; i++) {
let n = somBitrateNeurons[i];
if (n.bitrate < currentNeuron.bitrate && n.bitrate > maxSuitableBitrate && currentThroughput > n.bitrate) {
// possible downshiftable neuron
maxSuitableBitrate = n.bitrate;
result = n;
}
}
}
return result;
}
/**
*
* @param {object} mediaInfo
* @param {number} throughput
* @param {number} latency
* @param {number} bufferSize
* @param {number} playbackRate
* @param {number} currentQualityIndex
* @param {object} dynamicWeightsSelector
* @return {null|*}
*/
function getNextQuality(mediaInfo, throughput, latency, bufferSize, playbackRate, currentQualityIndex, dynamicWeightsSelector) {
// For Dynamic Weights Selector
let currentLatency = latency;
let currentBuffer = bufferSize;
let currentThroughput = throughput;
let somElements = _getSomBitrateNeurons(mediaInfo);
// normalize throughput
let throughputNormalized = throughput / bitrateNormalizationFactor;
// saturate values higher than 1
if (throughputNormalized > 1) {
throughputNormalized = _getMaxThroughput();
}
// normalize latency
latency = latency / latencyNormalizationFactor;
const targetLatency = 0;
const targetRebufferLevel = 0;
const targetSwitch = 0;
// 10K + video encoding is the recommended throughput
const throughputDelta = 10000;
logger.debug(`getNextQuality called throughput:${throughputNormalized} latency:${latency} bufferSize:${bufferSize} currentQualityIndex:${currentQualityIndex} playbackRate:${playbackRate}`);
let currentNeuron = somElements[currentQualityIndex];
let downloadTime = (currentNeuron.bitrate * dynamicWeightsSelector.getSegmentDuration()) / currentThroughput;
let rebuffer = Math.max(0, (downloadTime - currentBuffer));
// check buffer for possible stall
if (currentBuffer - downloadTime < dynamicWeightsSelector.getMinBuffer()) {
logger.debug(`Buffer is low for bitrate= ${currentNeuron.bitrate} downloadTime=${downloadTime} currentBuffer=${currentBuffer} rebuffer=${rebuffer}`);
return _getDownShiftNeuron(currentNeuron, currentThroughput).qualityIndex;
}
switch (weightSelectionMode) {
case WEIGHT_SELECTION_MODES.MANUAL:
_manualWeightSelection();
break;
case WEIGHT_SELECTION_MODES.RANDOM:
_randomWeightSelection(somElements);
break;
case WEIGHT_SELECTION_MODES.DYNAMIC:
_dynamicWeightSelection(dynamicWeightsSelector, somElements, currentLatency, currentBuffer, rebuffer, currentThroughput, playbackRate);
break;
default:
_dynamicWeightSelection(dynamicWeightsSelector, somElements, currentLatency, currentBuffer, rebuffer, currentThroughput, playbackRate);
}
let minDistance = null;
let minIndex = null;
let winnerNeuron = null;
for (let i = 0; i < somElements.length; i++) {
let somNeuron = somElements[i];
let somNeuronState = somNeuron.state;
let somData = [somNeuronState.throughput,
somNeuronState.latency,
somNeuronState.rebuffer,
somNeuronState.switch];
let distanceWeights = weights.slice();
let nextBuffer = dynamicWeightsSelector.getNextBufferWithBitrate(somNeuron.bitrate, currentBuffer, currentThroughput);
let isBufferLow = nextBuffer < dynamicWeightsSelector.getMinBuffer();
if (isBufferLow) {
logger.debug(`Buffer is low for bitrate=${somNeuron.bitrate} downloadTime=${downloadTime} currentBuffer=${currentBuffer} nextBuffer=${nextBuffer}`);
}
// special condition downshift immediately
if (somNeuron.bitrate > throughput - throughputDelta || isBufferLow) {
if (somNeuron.bitrate !== minBitrate) {
// encourage to pick smaller bitrates throughputWeight=100
distanceWeights[0] = 100;
}
}
// calculate the distance with the target
let distance = _getDistance(somData, [throughputNormalized, targetLatency, targetRebufferLevel, targetSwitch], distanceWeights);
if (minDistance === null || distance < minDistance) {
minDistance = distance;
minIndex = somNeuron.qualityIndex;
winnerNeuron = somNeuron;
}
}
// update current neuron and the neighbourhood with the calculated QoE
// will punish current if it is not picked
let bitrateSwitch = Math.abs(currentNeuron.bitrate - winnerNeuron.bitrate) / bitrateNormalizationFactor;
_updateNeurons(currentNeuron, somElements, [throughputNormalized, latency, rebuffer, bitrateSwitch]);
// update bmu and neighbours with targetQoE=1, targetLatency=0
_updateNeurons(winnerNeuron, somElements, [throughputNormalized, targetLatency, targetRebufferLevel, bitrateSwitch]);
return minIndex;
}
/**
* Option 1: Manual weights
* @private
*/
function _manualWeightSelection() {
let throughputWeight = 0.4;
let latencyWeight = 0.4;
let bufferWeight = 0.4;
let switchWeight = 0.4;
weights = [throughputWeight, latencyWeight, bufferWeight, switchWeight]; // throughput, latency, buffer, switch
}
/**
* Option 2: Random (Xavier) weights
* @param {array} somElements
* @private
*/
function _randomWeightSelection(somElements) {
weights = _getXavierWeights(somElements.length, 4);
}
/**
* Dynamic Weight Selector weights
* @param {object} dynamicWeightsSelector
* @param {array} somElements
* @param {number} currentLatency
* @param {number} currentBuffer
* @param {number} rebuffer
* @param {number} currentThroughput
* @param {number} playbackRate
* @private
*/
function _dynamicWeightSelection(dynamicWeightsSelector, somElements, currentLatency, currentBuffer, rebuffer, currentThroughput, playbackRate) {
if (!weights) {
weights = sortedCenters[sortedCenters.length - 1];
}
// Dynamic Weights Selector (step 2/2: find weights)
let weightVector = dynamicWeightsSelector.findWeightVector(somElements, currentLatency, currentBuffer, rebuffer, currentThroughput, playbackRate);
if (weightVector !== null && weightVector !== -1) { // null: something went wrong, -1: constraints not met
weights = weightVector;
}
}
/**
*
* @param {number }neuronCount
* @param {number }weightCount
* @return {array}
* @private
*/
function _getXavierWeights(neuronCount, weightCount) {
let W = [];
let upperBound = Math.sqrt((2 / neuronCount));
for (let i = 0; i < weightCount; i++) {
W.push(Math.random() * upperBound);
}
weights = W;
return weights;
}
/**
*
* @param {object} mediaInfo
* @return {array}
* @private
*/
function _getSomBitrateNeurons(mediaInfo) {
if (!somBitrateNeurons) {
somBitrateNeurons = [];
const bitrateList = mediaInfo.bitrateList;
let bitrateVector = [];
minBitrate = bitrateList[0].bandwidth;
bitrateList.forEach(element => {
bitrateVector.push(element.bandwidth);
if (element.bandwidth < minBitrate) {
minBitrate = element.bandwidth;
}
});
bitrateNormalizationFactor = _getMagnitude(bitrateVector);
for (let i = 0; i < bitrateList.length; i++) {
let neuron = {
qualityIndex: i,
bitrate: bitrateList[i].bandwidth,
state: {
// normalize throughputs
throughput: bitrateList[i].bandwidth / bitrateNormalizationFactor,
latency: 0,
rebuffer: 0,
switch: 0
}
};
somBitrateNeurons.push(neuron);
}
sortedCenters = _getInitialKmeansPlusPlusCenters(somBitrateNeurons);
}
return somBitrateNeurons;
}
/**
*
* @param {number} size
* @return {array}
* @private
*/
function _getRandomData(size) {
let dataArray = [];
for (let i = 0; i < size; i++) {
let data = [
Math.random() * _getMaxThroughput(), //throughput
Math.random(), //latency
Math.random(), //buffersize
Math.random() //switch
];
dataArray.push(data);
}
return dataArray;
}
/**
*
* @param {array} somElements
* @return {array}
* @private
*/
function _getInitialKmeansPlusPlusCenters(somElements) {
let centers = [];
let randomDataSet = _getRandomData(Math.pow(somElements.length, 2));
centers.push(randomDataSet[0]);
let distanceWeights = [1, 1, 1, 1];
for (let k = 1; k < somElements.length; k++) {
let nextPoint = null;
let maxDistance = null;
for (let i = 0; i < randomDataSet.length; i++) {
let currentPoint = randomDataSet[i];
let minDistance = null;
for (let j = 0; j < centers.length; j++) {
let distance = _getDistance(currentPoint, centers[j], distanceWeights);
if (minDistance === null || distance < minDistance) {
minDistance = distance;
}
}
if (maxDistance === null || minDistance > maxDistance) {
nextPoint = currentPoint;
maxDistance = minDistance;
}
}
centers.push(nextPoint);
}
// find the least similar center
let maxDistance = null;
let leastSimilarIndex = null;
for (let i = 0; i < centers.length; i++) {
let distance = 0;
for (let j = 0; j < centers.length; j++) {
if (i === j) continue;
distance += _getDistance(centers[i], centers[j], distanceWeights);
}
if (maxDistance === null || distance > maxDistance) {
maxDistance = distance;
leastSimilarIndex = i;
}
}
// move centers to sortedCenters
let sortedCenters = [];
sortedCenters.push(centers[leastSimilarIndex]);
centers.splice(leastSimilarIndex, 1);
while (centers.length > 0) {
let minDistance = null;
let minIndex = null;
for (let i = 0; i < centers.length; i++) {
let distance = _getDistance(sortedCenters[0], centers[i], distanceWeights);
if (minDistance === null || distance < minDistance) {
minDistance = distance;
minIndex = i;
}
}
sortedCenters.push(centers[minIndex]);
centers.splice(minIndex, 1);
}
return sortedCenters;
}
instance = {
getNextQuality,
reset
};
_setup();
return instance;
}
LearningAbrController.__dashjs_factory_name = 'LearningAbrController';
export default FactoryMaker.getClassFactory(LearningAbrController);