public int Act(double[] state)
{
var s = StateKey(state);
// act according to epsilon greedy policy
var a = 0;
var poss = AllowedActions(state);
var probs = new List <double>();
for (var i = 0; i < poss.Length; i++)
{
probs.Add(P[poss[i] * NS + s]);
}
// epsilon greedy policy
if (Tembo.Random() < Options.Epsilon)
{
a = poss[Tembo.RandomInt(0, poss.Length)]; // random available action
Explored = true;
}
else
{
a = poss[Tembo.SampleWeighted(probs.ToArray())];
Explored = false;
}
// shift state memory
s0 = s1;
a0 = a1;
s1 = s;
a1 = a;
return(a);
}
/// <summary>
/// Returns an action from a state
/// </summary>
/// <param name="state">state size must be equal to NumberOfStates</param>
/// <returns></returns>
public int Act(double[] state)
{
Tembo.Assert(state.Length == NumberOfStates, $"Current state({state.Length}) not equal to NS({NumberOfStates})");
var a = 0;
// convert to a Mat column vector
var s = new Matrix(NumberOfStates, 1);
s.Set(state);
// epsilon greedy policy
if (Tembo.Random() < Options.Epsilon)
{
a = Tembo.RandomInt(0, NumberOfActions);
}
else
{
// greedy wrt Q function
var amat = ForwardQ(Network, s, false);
a = Tembo.Maxi(amat.W); // returns index of argmax action
}
// shift state memory
this.s0 = this.s1;
this.a0 = this.a1;
this.s1 = s;
this.a1 = a;
return(a);
}
private void BLearning()
{
while (true)
{
if (Historical.Count < 20000)
{
//
Thread.Sleep(TimeSpan.FromMinutes(30));
}
var correct = 0.0;
var total = 0.0;
var options = new AgentOptions
{
Gamma = Tembo.Random(0.01, 0.99),
Epsilon = Tembo.Random(0.01, 0.75),
Alpha = Tembo.Random(0.01, 0.99),
ExperinceAddEvery = Tembo.RandomInt(1, 10000),
ExperienceSize = 0,
LearningSteps = Tembo.RandomInt(1, 10),
HiddenUnits = Tembo.RandomInt(100000, 100000000),
ErrorClamp = Tembo.Random(0.01, 1.0),
AdaptiveLearningSteps = true
};
var agent = new DQN(dqnAgent.NumberOfStates, dqnAgent.NumberOfActions, options);
for (var i = 0; i < Historical.Count; i++)
{
var spi = Historical.ElementAt(i);
var action = agent.Act(spi.Value.Values);
if (action == spi.Value.Output)
{
correct += 1;
agent.Learn(1);
}
else
{
agent.Learn(-1);
}
total += 1;
}
var winrate = (correct / total) * 100;
if (winrate > WinRate)
{
CN.Log($"NEW AGENT DISCOVERED --> WINRATE {winrate.ToString("p")}, CLASS: {AgentName}", 2);
Save();
dqnAgent = agent;
WinRate = winrate;
}
}
}