static void Main(string[] args)
{
Env env = new LunarLanderEnv();
var memory = new PriorityExperienceReplay(100000);
var qFunc = new DQN(env.ObservationSpace.Shape, env.ActionSpace.NumberOfValues(), new[] { 256, 128 }, 0.0001f, 0.999f, 32, memory)
{
MemoryInterval = 1,
EnableDoubleDQN = true,
TargetModelUpdateInterval = 4000,
TrainingEpochs = 1
};
Agent agent = new AgentQL("dqn_lunarlander", env, qFunc)
{
WarmupSteps = 5000,
MaxEpsilon = 1.0f,
MinEpsilon = 0.01f,
EpsilonDecay = 0.995f,
TrainInterval = 1,
RewardClipping = false,
TrainRenderInterval = 10,
Verbose = true,
RenderFreq = 80,
};
agent.Train(1500, 1500);
//agent.Load($"{agent.Name}_1500");
agent.Test(100, 400, 2);
}
static void Main(string[] args)
{
Env env = new CartPoleEnv();
//https://github.com/rlcode/reinforcement-learning/blob/master/2-cartpole/1-dqn/cartpole_dqn.py
var valueFunc = new DQN(env.ObservationSpace.Shape, env.ActionSpace.NumberOfValues(), new[] { 24, 24 }, 0.001f, 0.99f, 32, new ExperienceReplay(2000))
{
TargetModelUpdateInterval = 1
};
Agent agent = new AgentQL("dqn_cartpole", env, valueFunc)
{
Verbose = true,
RewardOnDone = -100,
EpsilonDecayMode = EEpsilonDecayMode.EveryStep,
EpsilonDecay = 0.999f,
WarmupSteps = 1000
};
agent.Train(300, 500);
Console.WriteLine($"Average reward {agent.Test(50, 300, 0)}");
//while (!env.Step((int)env.ActionSpace.Sample()[0], out var nextState, out var reward))
//{
// env.Render();
// Thread.Sleep(50);
//}
return;
}
static void Main(string[] args)
{
Env env = new TaxiEnv();
Agent agent = new AgentQL("qtable_taxi", env, new QTable(env.ObservationSpace.NumberOfValues(), env.ActionSpace.NumberOfValues(), 0.7f, 0.618f))
{
Verbose = true
};
agent.Train(50000, 100);
Console.WriteLine($"Average reward {agent.Test(100, 100, 1)}");
//while (!env.Step((int)env.ActionSpace.Sample()[0], out var nextState, out var reward))
// env.Render();
return;
}