private IEnumerator Start()
{
var settings = new DQNInitSettings()
{
};
double[] stateArr = new double[] { center.x - transform.position.x, center.y - transform.position.y };
var agent = new DQNAgent(settings, stateArr.Length, 4);
while (true)
{
int action = agent.Act(stateArr);
Vector2 movementVec = default;
switch (action)
{
case 0:
movementVec = new Vector2(1f, 0f) * Time.deltaTime * speed;
break;
case 1:
movementVec = new Vector2(-1f, 0f) * Time.deltaTime * speed;
break;
case 2:
movementVec = new Vector2(0f, 1f) * Time.deltaTime * speed;
break;
case 3:
movementVec = new Vector2(0f, -1f) * Time.deltaTime * speed;
break;
}
Vector2 pos = new Vector2(transform.position.x, transform.position.y);
if ((center - pos).sqrMagnitude > (center - pos + movementVec).sqrMagnitude)
{
Debug.Log("negative");
agent.Learn(-1.0);
}
else
{
Debug.Log("positive");
agent.Learn(1.0);
}
rgb.velocity = movementVec;
yield return(new WaitForFixedUpdate());
}
}
static void Main(string[] args)
{
var rnd = new Random();
int max = 10;
int min = 1;
int nextPrint = 0, act1 = 0, act0 = 0;
double total = 0, correct = 0;
var state = new[] { rnd.Next(min, max), rnd.Next(min, max), rnd.Next(min, max), rnd.Next(min, max) };
var opt = new TrainingOptions
{
Alpha = 0.001,
Epsilon = 0,
ErrorClamp = 0.002,
ExperienceAddEvery = 10,
ExperienceSize = 1000,
ExperienceStart = 0,
HiddenUnits = 5,
LearningSteps = 400
};
//we take 4 states i.e random numbers between 1 and 10
//we have 2 actions 1 if average of set is >5 and 0 if otherwise
//we reward agent with 1 for every correct and -1 otherwise
var agent = new DQNAgent(opt, state.Length, 2);
//how to properly use the DPAgent
//var agent2= new MyDPAgent();
//agent2.Reset(state.Length,2);
while (total < 50000)
{
state = new[] { rnd.Next(min, max), rnd.Next(min, max), rnd.Next(min, max), rnd.Next(min, max) };
var action = agent.Act(state);
if (action == 1)
{
act1++;
}
else
{
act0++;
}
if (state.Average() > 5 && action == 1)
{
agent.Learn(1);
correct++;
}
else if (state.Average() <= 5 && action == 0)
{
agent.Learn(1);
correct++;
}
else
{
agent.Learn(-1);
}
total++;
//nextPrint++;
if (total >= nextPrint)
{
Console.WriteLine("Score: " + (correct / total).ToString("P") + "Epoch: " + nextPrint);
Console.WriteLine("Action 1: " + act1 + " Action 0: " + act0);
nextPrint += 1000;
}
}
// Console.WriteLine("Score: " + (correct / total).ToString("P"));
Console.WriteLine("End");
File.AppendAllText(AppDomain.CurrentDomain.BaseDirectory + "DNQ.trr", agent.AgentToJson());
Console.ReadKey();
}
