# Evaluation

## Overview

Submissions are evaluated using **Root Mean Squared Error (RMSE)** between the predicted `settlement_index` and the actual observed values.

## Metric

The Root Mean Squared Error is calculated as:

$$\text{RMSE} = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2}$$

Where:

- $n$ is the number of planets in the test set
- $y_i$ is the actual settlement viability index for planet $i$
- $\hat{y}_i$ is your predicted settlement viability index for planet $i$

## Submission Format

For each `planet_name` in the test set, you must predict a continuous value for `settlement_index`. The target variable ranges from **0 to 100**, where higher values indicate planets more suitable for establishing successful, self-sustaining settlements.

Your submission file should contain a header and have the following format:

```
planet_name,settlement_index
Kepler-442c,78.45
Proxima Centauri d,34.21
Trappist-1f,62.89
...
```

## About the Target Variable

The `settlement_index` is a comprehensive metric that synthesizes multiple aspects of planetary assessment into a single predictive score. This index reflects the overall feasibility of establishing a permanent human presence on each exoplanet, considering factors that span:

- **Physical environment compatibility** with human physiology and existing technology
- **Resource availability** for sustaining long-term settlement
- **Safety considerations** that affect settler survival and mission success
- **Economic viability** including infrastructure costs and strategic importance
- **Long-term sustainability** factors for multi-generational habitation

The target represents expert evaluations from the interstellar exploration agency's settlement assessment protocols. Your challenge is to build a model that can accurately predict this viability score using the rich combination of structured planetary attributes, unstructured field reports, and environmental data provided for each candidate world.

## Evaluation Details

- **Lower RMSE is better** — a perfect score would be 0.0
- Predictions will be evaluated on a held-out test set of planets not included in the training data
- Your predictions should be continuous values; they will not be rounded before scoring
- Predictions outside the 0-100 range are permitted but may indicate model issues

## Tips

- The dataset includes both **structured tabular features** (atmospheric composition, orbital mechanics, resource indices) and **unstructured text fields** (exploration logs, environmental reports, incident reports)
- Consider leveraging both data types for optimal performance
- The `image_prompt` field contains rich descriptive information about each planet's surface conditions
- Pay attention to physical consistency between related variables
- Some planets represent edge cases with extreme or unusual characteristics

Good luck, and may your models identify the most promising new homes for humanity! 🚀