🚀 OptimRL: Group Relative Policy Optimization

OptimRL is a high-performance reinforcement learning library that introduces a groundbreaking algorithm, Group Relative Policy Optimization (GRPO). Designed to streamline the training of RL agents, GRPO eliminates the need for a critic network while ensuring robust performance with group-based advantage estimation and KL regularization. Whether you're building an AI to play games, optimize logistics, or manage resources, OptimRL provides state-of-the-art efficiency and stability.

🏅 Badges

🌟 Features

Why Choose OptimRL?

1. 🚫 Critic-Free Learning
   Traditional RL methods require training both an actor and a critic network. GRPO eliminates this dual-network requirement, cutting model complexity by 50% while retaining top-tier performance.

2. 👥 Group-Based Advantage Estimation
   GRPO introduces a novel way to normalize rewards within groups of experiences. This ensures:

   • Stable training across diverse reward scales.
   • Adaptive behavior for varying tasks and environments.

3. 📏 KL Regularization
   Prevent policy collapse with GRPO's built-in KL divergence regularization, ensuring:

   • Smoothed updates for policies.
   • Reliable and stable learning in any domain.

4. ⚡ Vectorized NumPy Operations with PyTorch Tensor Integration
   OptimRL leverages NumPy's vectorized operations and PyTorch's tensor computations with GPU acceleration for maximum performance. This hybrid implementation provides:

   • 10-100x speedups over pure Python through optimized array programming
   • Seamless CPU/GPU execution via PyTorch backend
   • Native integration with deep learning workflows
   • Full automatic differentiation support

5. 🔄 Experience Replay Buffer
   Improve sample efficiency with built-in experience replay:

   • Learn from past experiences multiple times
   • Reduce correlation between consecutive samples
   • Configurable buffer capacity and batch sizes

6. 🔄 Continuous Action Space Support
   Train agents in environments with continuous control:

   • Gaussian policy implementation for continuous actions
   • Configurable action bounds
   • Adaptive standard deviation for exploration

---

🛠️ Installation

For End Users

Simply install from PyPI:

pip install optimrl

For Developers

Clone the repository and set up a development environment:

git clone https://github.com/subaashnair/optimrl.git
cd optimrl
pip install -e '.[dev]'

---

⚡ Quick Start

Discrete Action Space Example (CartPole)

import torch
import torch.nn as nn
import torch.optim as optim
import gym
from optimrl import create_agent

# Define a simple policy network
class PolicyNetwork(nn.Module):
    def __init__(self, input_dim, output_dim):
        super().__init__()
        self.network = nn.Sequential(
            nn.Linear(input_dim, 64),
            nn.ReLU(),
            nn.Linear(64, output_dim),
            nn.LogSoftmax(dim=-1)
        )
        
    def forward(self, x):
        return self.network(x)

# Create environment and network
env = gym.make('CartPole-v1')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.n
policy = PolicyNetwork(state_dim, action_dim)

# Create GRPO agent
agent = create_agent(
    "grpo",
    policy_network=policy,
    optimizer_class=optim.Adam,
    learning_rate=0.001,
    gamma=0.99,
    grpo_params={"epsilon": 0.2, "beta": 0.01},
    buffer_capacity=10000,
    batch_size=32
)

# Training loop
state, _ = env.reset()
for step in range(1000):
    action = agent.act(state)
    next_state, reward, done, truncated, _ = env.step(action)
    agent.store_experience(reward, done)
    
    if done or truncated:
        state, _ = env.reset()
        agent.update()  # Update policy after episode ends
    else:
        state = next_state

Complete CartPole Implementation

For a complete implementation of CartPole with OptimRL, check out our examples in the simple_test directory:

• cartpole_simple.py: Basic implementation with GRPO
• cartpole_improved.py: Improved implementation with tuned parameters
• cartpole_final.py: Final implementation with optimized performance
• cartpole_tuned.py: Enhanced implementation with advanced features
• cartpole_simple_pg.py: Vanilla Policy Gradient implementation for comparison

The vanilla policy gradient implementation (cartpole_simple_pg.py) achieves excellent performance on CartPole-v1, reaching the maximum reward of 500 consistently. It serves as a useful baseline for comparing against the GRPO implementations.

Continuous Action Space Example (Pendulum)

import torch
import torch.nn as nn
import torch.optim as optim
import gym
from optimrl import create_agent

# Define a continuous policy network
class ContinuousPolicyNetwork(nn.Module):
    def __init__(self, input_dim, action_dim):
        super().__init__()
        self.shared_layers = nn.Sequential(
            nn.Linear(input_dim, 64),
            nn.ReLU(),
            nn.Linear(64, 64),
            nn.ReLU()
        )
        # Output both mean and log_std for each action dimension
        self.output_layer = nn.Linear(64, action_dim * 2)
        
    def forward(self, x):
        x = self.shared_layers(x)
        return self.output_layer(x)

# Create environment and network
env = gym.make('Pendulum-v1')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bounds = (env.action_space.low[0], env.action_space.high[0])
policy = ContinuousPolicyNetwork(state_dim, action_dim)

# Create Continuous GRPO agent
agent = create_agent(
    "continuous_grpo",
    policy_network=policy,
    optimizer_class=optim.Adam,
    action_dim=action_dim,
    learning_rate=0.0005,
    gamma=0.99,
    grpo_params={"epsilon": 0.2, "beta": 0.01},
    buffer_capacity=10000,
    batch_size=64,
    min_std=0.01,
    action_bounds=action_bounds
)

# Training loop
state, _ = env.reset()
for step in range(1000):
    action = agent.act(state)
    next_state, reward, done, truncated, _ = env.step(action)
    agent.store_experience(reward, done)
    
    if done or truncated:
        state, _ = env.reset()
        agent.update()  # Update policy after episode ends
    else:
        state = next_state

📊 Performance Comparison

Our simple policy gradient implementation consistently solves the CartPole-v1 environment in under 1000 episodes, achieving the maximum reward of 500. The GRPO implementations offer competitive performance with additional benefits:

• Lower variance: More stable learning across different random seeds
• Improved sample efficiency: Learns from fewer interactions with the environment
• Better regularization: Prevents policy collapse during training

Kaggle Notebook

You can view the "OptimRL Trading Experiment" notebook on Kaggle:

Alternatively, you can open the notebook locally as an .ipynb file: Open the OptimRL Trading Experiment Notebook (.ipynb)

---

🤝 Contributing

We're excited to have you onboard! Here's how you can help improve OptimRL:

1. Fork the repo.
2. Create a feature branch:

   git checkout -b feature/AmazingFeature
   

3. Commit your changes:

   git commit -m 'Add some AmazingFeature'
   

4. Push to the branch:

   git push origin feature/AmazingFeature
   

5. Open a Pull Request.

Before submitting, make sure you run all tests:

pytest tests/

---

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

📚 Citation

If you use OptimRL in your research, please cite:

@software{optimrl2024,
  title={OptimRL: Group Relative Policy Optimization},
  author={Subashan Nair},
  year={2024},
  url={https://github.com/subaashnair/optimrl}
}

---