public void TestDSLocalConcurrent()
{
const float Eps = 0.2f;
string vwArgs = "--cb_explore_adf --epsilon " + Eps.ToString();
DecisionServiceLocal <SimpleADFContext> dsLocal = new DecisionServiceLocal <SimpleADFContext>(vwArgs, 1, TimeSpan.MaxValue);
var context = new SimpleADFContext {
Id = "Shared", Actions = new int[] { 1, 2, 3 }
};
const int NumThreads = 16;
const int NumEventsPerThread = 25;
List <Thread> threads = new List <Thread>(NumThreads);
int[] targetActionCnts = Enumerable.Repeat <int>(0, NumThreads).ToArray();
int idCounter = 0;
for (int i = 0; i < NumThreads; i++)
{
threads.Add(new Thread(() =>
{
int id = Interlocked.Increment(ref idCounter) - 1;
Console.WriteLine("in thread {0}", id);
int action;
for (int j = 0; j < NumEventsPerThread; j++)
{
string guid = Guid.NewGuid().ToString();
action = dsLocal.ChooseActionAsync(guid, context, 1).Result;
dsLocal.ReportRewardAndComplete((action == 2) ? 1.0f : 0.0f, guid);
targetActionCnts[id] += (action == 2) ? 1 : 0;
}
}));
}
foreach (Thread t in threads)
{
t.Start();
}
foreach (Thread t in threads)
{
t.Join();
}
// Since the model is updated after each datapoint, we expect most exploit predictions
// (1 - Eps) to be the middle action, but allow fro some slack.
Console.WriteLine("Sum of target is {0}, total is {1}", targetActionCnts.Sum(), NumThreads * NumEventsPerThread);
Assert.IsTrue(targetActionCnts.Sum() * 1.0 / (NumThreads * NumEventsPerThread) >= (1 - Eps) * 0.9);
}
public void TestDSLocalModelLearning()
{
const int NumEvents = 100;
const float Eps = 0.2f;
string vwArgs = "--cb_explore_adf --epsilon " + Eps.ToString();
// Test both generic class and json string typed versions of DS local
DecisionServiceLocal <SimpleADFContext> dsLocal = new DecisionServiceLocal <SimpleADFContext>(vwArgs, 1, TimeSpan.MaxValue);
DecisionServiceLocal <string> dsLocalJson = new DecisionServiceLocal <string>(vwArgs, 1, TimeSpan.MaxValue);
var context = new SimpleADFContext {
Id = "Shared", Actions = new int[] { 1, 2, 3 }
};
int action, actionJson;
int targetActionCnt = 0, targetActionJsonCnt = 0;
// Generate interactions and reward the model for the middle action only (learning the
// lowest/highest can be done even with bad featurization, which we want to catch).
for (int i = 0; i < NumEvents; i++)
{
string guid = Guid.NewGuid().ToString();
// Test generic class type
action = dsLocal.ChooseActionAsync(guid, context, 1).Result;
dsLocal.ReportRewardAndComplete((action == 2) ? 1.0f : 0.0f, guid);
targetActionCnt += (action == 2) ? 1 : 0;
string contextJson = JsonConvert.SerializeObject(context);
actionJson = dsLocalJson.ChooseActionAsync(guid, contextJson, 1).Result;
//TODO: The examples should look identical to VW, so predictions should be identical
//Assert.IsTrue(action == actionJson);
dsLocalJson.ReportRewardAndComplete((actionJson == 2) ? 1.0f : 0.0f, guid);
targetActionJsonCnt += (actionJson == 2) ? 1 : 0;
}
// Since the model is updated after each datapoint, we expect most exploit predictions
// (1 - Eps) to be the middle action, but allow fro some slack.
Assert.IsTrue(targetActionCnt * 1.0 / NumEvents >= (1 - Eps) * 0.9);
Assert.IsTrue(targetActionJsonCnt * 1.0 / NumEvents >= (1 - Eps) * 0.9);
}
public static IReadOnlyCollection <SimpleActionFeature> GetFeaturesFromContext(SimpleADFContext context)
{
return(context.ActionDependentFeatures);
}