private int SendEvents(DecisionServiceDeployment deployment, WebClient client, int featureIndex, bool sendReward = true)
{
const float reward = 2.0F;
var expectedEvents = 0;
string contextString = $"{{a: \"{features[featureIndex]}\"}}";
var responseJObj = InteractionParts1and2(deployment, contextType, contextString);
int action = (int)responseJObj["Action"];
// Feature | Action
// A --> 1
// B --> 2
// C --> 3
// D --> 4
// only report in 50% of the cases
if (sendReward && rnd.NextDouble() < .75 && action - 1 == featureIndex)
{
InteractionPart3(deployment, responseJObj, reward);
expectedEvents = 1;
}
var stat = string.Format("'{0}' '{1}' ", features[featureIndex], action);
int count;
if (freq.TryGetValue(stat, out count))
{
freq[stat]++;
}
else
{
freq.Add(stat, count);
}
return(expectedEvents);
}
public JObject InteractionParts1and2(DecisionServiceDeployment deployment, string contextType, string contextString)
{
string contextUri = string.Format(CultureInfo.InvariantCulture, "{0}/API/{1}", deployment.ManagementCenterUrl, contextType);
byte[] context = System.Text.Encoding.ASCII.GetBytes(contextString);
var response = wc.UploadData(contextUri, "POST", context);
var utf8response = UnicodeEncoding.UTF8.GetString(response);
JObject responseJObj = JObject.Parse(utf8response);
return(responseJObj);
}
public string InteractionPart3(DecisionServiceDeployment deployment, JObject responseJObj, float reward)
{
string eventID = (string)responseJObj["EventId"];
string rewardUri = string.Format(CultureInfo.InvariantCulture, "{0}/API/reward/?eventId={1}", deployment.ManagementCenterUrl, eventID);
string rewardString = reward.ToString();
byte[] rewardBytes = System.Text.Encoding.ASCII.GetBytes(rewardString);
var response = wc.UploadData(rewardUri, "POST", rewardBytes);
string utf8response = UnicodeEncoding.UTF8.GetString(response);
return(utf8response);
}
public void E2ERankerStochasticRewards(DecisionServiceDeployment deployment)
{
// Create configuration for the decision service
float initialEpsilon = .5f;
deployment.ConfigureDecisionService(trainArguments: "--cb_explore_adf --cb_type dr -q :: --epsilon 0.2", initialExplorationEpsilon: initialEpsilon);
string settingsBlobUri = deployment.SettingsUrl;
deployment.OnlineTrainerWaitForStartup();
float percentCorrect = UploadFoodContextData(deployment, settingsBlobUri, firstPass: true);
Assert.IsTrue(percentCorrect < initialEpsilon);
percentCorrect = UploadFoodContextData(deployment, settingsBlobUri, firstPass: false);
Assert.IsTrue(percentCorrect > .8f);
}
public async Task SimplePolicyTest(DecisionServiceDeployment deployment)
{
deployment.OnlineTrainerWaitForStartup();
deployment.ConfigureDecisionService("--cb_explore 4 --epsilon 0", initialExplorationEpsilon: 1, isExplorationEnabled: true);
// 4 Actions
// why does this need to be different from default?
var config = new DecisionServiceConfiguration(deployment.SettingsUrl)
{
InteractionUploadConfiguration = new BatchingConfiguration
{
MaxEventCount = 64
},
ObservationUploadConfiguration = new BatchingConfiguration
{
MaxEventCount = 64
},
PollingForModelPeriod = TimeSpan.FromMinutes(5)
};
config.InteractionUploadConfiguration.ErrorHandler += JoinServiceBatchConfiguration_ErrorHandler;
config.InteractionUploadConfiguration.SuccessHandler += JoinServiceBatchConfiguration_SuccessHandler;
this.features = new string[] { "a", "b", "c", "d" };
this.freq = new Dictionary <string, int>();
this.rnd = new Random(123);
deployment.OnlineTrainerReset();
{
var expectedEvents = 0;
using (var client = Microsoft.Research.MultiWorldTesting.ClientLibrary.DecisionService.Create <MyContext>(config))
{
for (int i = 0; i < 50; i++)
{
expectedEvents += SendEvents(client, 128);
// Thread.Sleep(500);
}
}
// TODO: flush doesn't work
// Assert.AreEqual(expectedEvents, this.eventCount);
}
// 4 actions times 4 feature values
Assert.AreEqual(4 * 4, freq.Keys.Count);
Console.WriteLine("Exploration");
var total = freq.Values.Sum();
foreach (var k in freq.Keys.OrderBy(k => k))
{
var f = freq[k] / (float)total;
Assert.IsTrue(f < 0.08);
Console.WriteLine("{0} | {1}", k, f);
}
freq.Clear();
await Task.Delay(TimeSpan.FromMinutes(2));
// TODO: update eps: 0
using (var client = Microsoft.Research.MultiWorldTesting.ClientLibrary.DecisionService.Create <MyContext>(config))
{
int i;
for (i = 0; i < 120; i++)
{
try
{
client.DownloadModelAndUpdate(new System.Threading.CancellationToken()).Wait();
break;
}
catch (Exception)
{
await Task.Delay(TimeSpan.FromSeconds(1));
}
}
Assert.IsTrue(i < 30, "Unable to download model");
for (i = 0; i < 1024; i++)
{
var key = Guid.NewGuid().ToString();
var featureIndex = i % features.Length;
var action = client.ChooseAction(key, new MyContext {
Feature = features[featureIndex]
});
var stat = string.Format("'{0}' '{1}' ", features[featureIndex], action);
int count;
if (freq.TryGetValue(stat, out count))
{
freq[stat]++;
}
else
{
freq.Add(stat, count);
}
}
}
Console.WriteLine("Exploitation");
total = freq.Values.Sum();
foreach (var k in freq.Keys.OrderBy(k => k))
{
var f = freq[k] / (float)total;
Assert.AreEqual(0.25f, f, 0.1);
Console.WriteLine("{0} | {1}", k, f);
}
}
private float UploadFoodContextData(DecisionServiceDeployment deployment, string settingsBlobUri, bool firstPass)
{
var serviceConfig = new DecisionServiceConfiguration(settingsBlobUri);
if (firstPass)
{
serviceConfig.PollingForModelPeriod = TimeSpan.MinValue;
deployment.OnlineTrainerReset();
}
using (var service = DecisionService.Create <FoodContext>(serviceConfig))
{
if (!firstPass)
{
Thread.Sleep(10000);
}
string uniqueKey = "scratch-key-gal";
string[] locations = { "HealthyTown", "LessHealthyTown" };
var rg = new Random(uniqueKey.GetHashCode());
int numActions = 3; // ["Hamburger deal 1", "Hamburger deal 2" (better), "Salad deal"]
var csv = new StringBuilder();
int counterCorrect = 0;
int counterTotal = 0;
var header = "Location,Action,Reward";
csv.AppendLine(header);
// number of iterations
for (int i = 0; i < 10000 * locations.Length; i++)
{
// randomly select a location
int iL = rg.Next(0, locations.Length);
string location = locations[iL];
DateTime timeStamp = DateTime.UtcNow;
string key = uniqueKey + Guid.NewGuid().ToString();
FoodContext currentContext = new FoodContext();
currentContext.UserLocation = location;
currentContext.Actions = Enumerable.Range(1, numActions).ToArray();
int[] action = service.ChooseRanking(key, currentContext);
counterTotal += 1;
// We expect healthy town to get salad and unhealthy town to get the second burger (action 2)
if (location.Equals("HealthyTown") && action[0] == 3)
{
counterCorrect += 1;
}
else if (location.Equals("LessHealthyTown") && action[0] == 2)
{
counterCorrect += 1;
}
var csvLocation = location;
var csvAction = action[0].ToString();
float reward = 0;
double currentRand = rg.NextDouble();
if (location.Equals("HealthyTown"))
{
// for healthy town, buy burger 1 with probability 0.1, burger 2 with probability 0.15, salad with probability 0.6
if ((action[0] == 1 && currentRand < 0.1) ||
(action[0] == 2 && currentRand < 0.15) ||
(action[0] == 3 && currentRand < 0.6))
{
reward = 10;
}
}
else
{
// for unhealthy town, buy burger 1 with probability 0.4, burger 2 with probability 0.6, salad with probability 0.2
if ((action[0] == 1 && currentRand < 0.4) ||
(action[0] == 2 && currentRand < 0.6) ||
(action[0] == 3 && currentRand < 0.2))
{
reward = 10;
}
}
service.ReportReward(reward, key);
var newLine = string.Format("{0},{1},{2}", csvLocation, csvAction, "0");
csv.AppendLine(newLine);
System.Threading.Thread.Sleep(1);
}
return((float)counterCorrect / counterTotal);
}
}