asieradzk / rl_matrix

Reinforcement Learning Agents in .NET

🚀 RL Matrix - Pure C# Deep Reinforcement Learning Experience with TorchSharp!

Dive into the future of type-safe Deep Reinforcement Learning with .NET & RL Matrix, powered by the might of TorchSharp. RL Matrix stands out as a user-friendly toolkit offering a collection of RL algorithms—primed for plug, play, and prosper!

🌟 Featured Algorithms:

• PPO
• DQN
• Both have 1D (Feed forward) and 2D (CNN) variants
• 0.1.2 Adds multi-head continous (PPO) discrete (PPO, DQN) and mixed (PPO) actions. See IEnvironment and IContinousEnvironment.
• 0.2.0 Adds working-ish PPO GAIL. And overhauls training method for stepwise
• 0.2.0 Adds multi-environment training
• 0.2.0 Includes Godot examples and RLMatrix.Godot nuget package for easy setup
• Only tested single-head discrete output so please open issue if it doesnt work.

🎯 What Sparks RL Matrix?

While embarking on my RL journey, I sensed a gap in the reinforcement learning world even with TorchSharp's solid foundation. It struck me—C# is the ideal choice for RL outside research circles, thanks to its pristine and intuitive coding experience. No more guessing games in environment or agent building!

With RL Matrix, our vision is to offer a seamless experience. By simply incorporating the IEnvironment interface, you're equipped to rapidly craft and unleash Reinforcement Learning Agents (RL Agents). Switching between algorithms? A breeze! It’s our nod to the elegance of Matlab's toolkit methodology.

🚀 Getting Started:

Peek into the /examples/ directory for illustrative code. But to give you a quick start:

1. Craft an IEnvironment class: Comply with reinforcement learning guidelines, defining your observation shapes and action count:

public class CartPole : IEnvironment<float[]>
{
  public int stepCounter { get; set; }
  public int maxSteps { get; set; }
  public bool isDone { get; set; }
  public OneOf<int, (int, int)> stateSize { get; set; }
  public int actionSize { get; set; }

  CartPoleEnv myEnv;

  private float[] myState;

  public CartPole()
  {
      Initialise();
  }

  public float[] GetCurrentState()
  {
      if (myState == null)
          myState = new float[4] {0,0,0,0};
      return myState;
  }

  public void Initialise()
  {
      myEnv = new CartPoleEnv(WinFormEnvViewer.Factory);
      stepCounter = 0;
      maxSteps = 100000;
      stateSize = myEnv.ObservationSpace.Shape.Size;
      actionSize = myEnv.ActionSpace.Shape.Size;
      myEnv.Reset();
      isDone = false; 
  }

  public void Reset()
  {
  //For instance:
      myEnv.Reset();
      isDone = false;
      stepCounter = 0;
  }

  public float Step(int actionId)
  {
  //Whatever step logic, returns reward
      return reward;
  }
}

2. Agent Instance & Training: Spawn an agent for your environment and ignite the Step method:

var opts = new DQNAgentOptions(batchSize: 64, memorySize: 10000, gamma: 0.99f, epsStart: 1f, epsEnd: 0.05f, epsDecay: 50f, tau: 0.005f, lr: 1e-4f, displayPlot: myChart);
var env = new List<IEnvironment<float[]>> { new CartPole(), new CartPole() };
var myAgent = new DQNAgent<float[]>(opts, env);

for (int i = 0; i < 10000; i++)
{
    myAgent.Step();
}

Notice that TrainEpisode method was removed.

📌 Current Roadmap:

-Add RNN support for PPO and DQN -Add variations for multi-head output for PPO and DQN -More Godot examples testing multi-head continous+discrete action spaces -Create Godot plugin -Fully develop workflow for Gail and imitation learning As we innovate, anticipate breaking changes. We'll keep you in the loop!

💌 Get in Touch:

Questions? Ideas? Collaborations? Drop a line at: 📧 [email protected]

🤝 Join the Journey:

We believe in collective brilliance! If you're inspired to enhance RL Matrix, please open an issue or PR. Let’s shape the future of RL together!

rl_matrix's Issues

Reinforcement Learning The Identity Function

Context

I was having difficulty using reinforcement learning for more complex problems, so I made a simple project to test the learning potential. In the code below, I am trying to teach the Boolean identity function to the agent, but the agent gives random answers before and after training.

Code

using OneOf;
using RLMatrix; //NuGet v0.2.0

namespace ReinforcementLearning.Trivial;

internal class Program
{
    static void Main(string[] args)
    {
        var options = new PPOAgentOptions(lr: .005f);
        List<IEnvironment<float[]>> environments = [new TrivialEnvironment(), new TrivialEnvironment(), new TrivialEnvironment()];
        var agent = new PPOAgent<float[]>(options, environments, new PPONetProviderBase<float[]>(16, 2));

        Console.WriteLine($"Pre-training: {Evaluate(agent)}");

        for (int i = 0; i < 10000; i++)
        {
            Console.Write(i);
            agent.Step();
            Console.Write("\b\b\b\b\b\b\b\b\b\b\b\b");
        }

        Console.WriteLine($"Post-training: {Evaluate(agent)}");
    }

    private static float Evaluate(PPOAgent<float[]> agent, int count = 1000)
    {
        float[] state = new float[1];
        int correct = 0;
        for (int i = 0; i < count; i++)
        {
            state[0] = TrivialEnvironment.RandomValue();
            (int[] greedyDiscreteActions, float[] meanContinuousActions) = agent.SelectAction(state, false);
            if (state[0] == greedyDiscreteActions[0])
            {
                correct++;
            }
        }
        return (float)correct / count;
    }

    private sealed class TrivialEnvironment : IEnvironment<float[]>
    {
        public const float CorrectAnswerReward = 1;
        public const float WrongAnswerPenalty = -1;

        private readonly float[] state = [RandomValue()];

        public static int RandomValue()
        {
            return Random.Shared.Next(2);
        }

        public int stepCounter { get; set; }
        public int maxSteps { get; set; } = int.MinValue;
        public bool isDone { get; set; }
        public OneOf<int, (int, int)> stateSize { get; set; } = 1;
        public int[] actionSize { get; set; } = [2];

        public float[] GetCurrentState() => state;

        public TrivialEnvironment() => Initialise();

        public void Initialise() => Reset();

        public void Reset()
        {
            stepCounter = 0;
            isDone = false;
        }

        public float Step(int[] actionsIds)
        {
            float input = state[0];
            float output = actionsIds[0];

            state[0] = RandomValue();

            return input == output ? CorrectAnswerReward : WrongAnswerPenalty;
        }
    }
}

Am I using the library incorrectly?

Train one observation at a time

Thank you for the RLMatrix.

I have a complex environment that cannot be simulated, in which interdependent observations are generated one at a time, so running training on an entire episode at once is not an option. I had to add a TrainObservation() based on your TrainEpisode() code. Please add this minor fix.

Thanks

Introduce Test Driven Development!

It is now possible to do Test Driven Development using VS2022 to create Gym 2D and 3D RL environments for testing

using the latest gdUnit4API TestAdaptor on VS2022

Create Github Discussion

This will help the project discussion