public void GetQueryable()
{
var service = new MemoryCollectionFactory();
var result1 = service.GetOrCreateCollection <string>();
result1.Add("Foo");
var result2 = service.GetOrCreateCollection <string>();
result2.Add("Bar");
Assert.Equal(result1.Count, result2.Count);
}
/// <summary>
/// The Run method provides the main loop that performs the following steps:
/// 1.) get state
/// 2.) build experience
/// 3.) create policy gradients
/// 4.) train on experiences
/// </summary>
/// <param name="phase">Specifies the phae.</param>
/// <param name="nN">Specifies the number of iterations (based on the ITERATION_TYPE) to run, or -1 to ignore.</param>
/// <param name="type">Specifies the iteration type (default = ITERATION).</param>
/// <param name="step">Specifies the training step to take, if any. This is only used when debugging.</param>
public void Run(Phase phase, int nN, ITERATOR_TYPE type, TRAIN_STEP step)
{
IMemoryCollection iMemory = MemoryCollectionFactory.CreateMemory(m_memType, m_nMemorySize, m_fPriorityAlpha);
int nIteration = 1;
double dfRunningReward = 0;
double dfEpisodeReward = 0;
int nEpisode = 0;
bool bDifferent = false;
StateBase state = getData(phase, -1, -1);
// Preprocess the observation.
SimpleDatum x = m_brain.Preprocess(state, m_bUseRawInput, out bDifferent, true);
// Set the initial target model to the current model.
m_brain.UpdateTargetModel();
while (!m_brain.Cancel.WaitOne(0) && !isAtIteration(nN, type, nIteration, nEpisode))
{
// Forward the policy network and sample an action.
int action = m_brain.act(x, state.Clip, state.ActionCount);
// Take the next step using the action
StateBase state_next = getData(phase, action, nIteration);
// Preprocess the next observation.
SimpleDatum x_next = m_brain.Preprocess(state_next, m_bUseRawInput, out bDifferent);
if (!bDifferent)
{
m_brain.Log.WriteLine("WARNING: The current state is the same as the previous state!");
}
// Build up episode memory, using reward for taking the action.
iMemory.Add(new MemoryItem(state, x, action, state_next, x_next, state_next.Reward, state_next.Done, nIteration, nEpisode));
dfEpisodeReward += state_next.Reward;
// Do the training
if (iMemory.Count > m_brain.BatchSize)
{
double dfBeta = beta_by_frame(nIteration);
MemoryCollection rgSamples = iMemory.GetSamples(m_random, m_brain.BatchSize, dfBeta);
m_brain.Train(nIteration, rgSamples, state.ActionCount);
iMemory.Update(rgSamples);
if (nIteration % m_nUpdateTargetFreq == 0)
{
m_brain.UpdateTargetModel();
}
}
if (state_next.Done)
{
// Update reward running
dfRunningReward = dfRunningReward * 0.99 + dfEpisodeReward * 0.01;
nEpisode++;
updateStatus(nIteration, nEpisode, dfEpisodeReward, dfRunningReward, 0, 0, m_brain.GetModelUpdated());
state = getData(phase, -1, -1);
x = m_brain.Preprocess(state, m_bUseRawInput, out bDifferent, true);
dfEpisodeReward = 0;
}
else
{
state = state_next;
x = x_next;
}
nIteration++;
}
iMemory.CleanUp();
}
