README.md

---
language: en
tags:
- evolutionary-strategy
- cma-es
- gymnasium
- cartpole
- optimization
library_name: custom
datasets:
- gymnasium/CartPole-v1
metrics:
- mean_episode_length
model-index:
- name: CartPole
  results:
  - task:
      type: optimization
      name: CartPole-v1
    dataset:
      name: gymnasium/CartPole-v1
      type: gymnasium
    metrics:
    - type: mean_episode_length
      value: 500
      name: Mean Episode Length
license: mit
pipeline_tag: reinforcement-learning
---

# CartPole Solution

This model solves CartPole-v1 using CMA-ES with a linear policy.

### Training Convergence
![Training Convergence](assets/training_convergence.png)
*Figure: Training convergence showing the mean fitness (episode length) across generations. The model achieves optimal performance (500 steps) within 3 generations.*

## Model Details

### Model Description

This is a linear policy model for the CartPole-v1 environment that:
- Uses a simple weight matrix to map 4D state inputs to 2D action outputs
- Achieves optimal performance (500/500 steps) consistently
- Was optimized using CMA-ES, requiring only 3 generations for convergence
- Demonstrates sample-efficient learning for the CartPole balancing task

```python
def get_action(self, observation):
    observation = np.array(observation, dtype=np.float32)
    action_scores = np.dot(observation, self.weights)
    return int(np.argmax(action_scores))
```

- **Developed by:** Niladri Das
- **Model type:** Linear Policy
- **Language:** Python
- **License:** MIT
- **Finetuned from model:** No (trained from scratch)

### Model Sources

- **Hugging Face:** https://huggingface.co/harpertoken/harpertoken-cartpole



## Uses

### Direct Use

The model is designed for:
1. Solving the CartPole-v1 environment from Gymnasium
2. Demonstrating CMA-ES optimization for RL tasks
3. Serving as a baseline for comparison with other algorithms
4. Educational purposes in evolutionary strategies

### Out-of-Scope Use

The model should not be used for:
1. Complex control tasks beyond CartPole
2. Real-world robotics applications
3. Tasks requiring non-linear policies
4. Environments with partial observability

## Bias, Risks, and Limitations

### Technical Limitations
- Limited to CartPole-v1 environment
- Requires full state observation
- Linear policy architecture
- No transfer learning capability
- Environment-specific solution

### Performance Limitations
- May not handle significant environment variations
- No adaptation to changing dynamics
- Limited by linear policy capacity
- Requires precise state information

### Recommendations

Users should:
1. Only use for CartPole-v1 environment
2. Ensure full state observability
3. Understand the limitations of linear policies
4. Consider more complex architectures for other tasks
5. Validate performance in their specific setup

## How to Get Started with the Model

### Method 1: Using the CMAESAgent Class

```python
from model import CMAESAgent

# Load the model
agent = CMAESAgent.from_pretrained("harpertoken/harpertoken-cartpole")

# Evaluate
mean_reward, std_reward = agent.evaluate(num_episodes=5)
print(f"Mean reward: {mean_reward:.2f} ± {std_reward:.2f}")
```

### Method 2: Manual Implementation

```python
import numpy as np
from gymnasium import make

# Load model weights
weights = np.load('model_weights.npy')  # 4x2 matrix

# Create environment
env = make('CartPole-v1')

# Run inference
def get_action(observation):
    logits = observation @ weights
    return int(np.argmax(logits))

observation, _ = env.reset()
while True:
    action = get_action(observation)
    observation, reward, done, truncated, info = env.step(action)
    if done or truncated:
        break
```

## Training Details

### Training Data

- **Environment:** Gymnasium CartPole-v1
- **State Space:** 4D continuous (cart position, velocity, pole angle, angular velocity)
- **Action Space:** 2D discrete (left, right)
- **Reward:** +1 for each step, max 500 steps
- **Episode Termination:** Pole angle > 15°, cart position > 2.4, or 500 steps reached
- **Training Approach:** Direct environment interaction (no pre-collected dataset)

### Training Procedure

#### Training Hyperparameters

- **Algorithm:** CMA-ES
- **Population size:** 16
- **Number of generations:** 100 (early convergence by generation 3)
- **Initial step size:** 0.5
- **Parameters:** 8 (4x2 weight matrix)
- **Training regime:** Single precision (fp32)

#### Hardware Requirements

- **CPU:** Single core sufficient
- **Memory:** <100MB RAM
- **GPU:** Not required
- **Training time:** ~5 minutes on standard CPU

### Evaluation

#### Testing Data & Metrics

- **Environment:** Same as training (CartPole-v1)
- **Episodes:** 100 test episodes
- **Metrics:** Episode length, success rate

#### Results

- **Average Episode Length:** 500.0 ±0.0
- **Success Rate:** 100%
- **Convergence:** Achieved in 3 generations
- **Final Population Mean:** 500.00
- **Best Performance:** 500/500 consistently

## Implementation Details

The implementation employs a straightforward linear policy:

```python
class CMAESAgent:
    def __init__(self, env_name):
        self.env = gym.make(env_name)
        self.observation_space = self.env.observation_space.shape[0]  # 4 for CartPole
        self.action_space = self.env.action_space.n  # 2 for CartPole
        self.num_params = self.observation_space * self.action_space  # 8 total parameters
        self.weights = None
    
    def get_action(self, observation):
        observation = np.array(observation, dtype=np.float32)
        action_scores = np.dot(observation, self.weights)
        return int(np.argmax(action_scores))
```

The model's simplicity demonstrates that CartPole's optimal control policy is approximately linear in the state variables.

## Environmental Impact

- **Training time:** ~5 minutes
- **Hardware:** Standard CPU
- **Energy consumption:** Negligible (<0.001 kWh)
- **CO2 emissions:** Minimal (<0.001 kg)

## Citation

**BibTeX:**
```bibtex
@misc{das2025cartpole,
  author = {Niladri Das},
  title = {CartPole Solution},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
  howpublished = {https://huggingface.co/harpertoken/harpertoken-cartpole}
}