🌌 tileX — Genetic Algorithm Scheduler and NN Surrogate

This code, so far in a notebook form, implements and experiments with a Genetic Algorithm (GA) based sky observation scheduler and a neural-network surrogate model for fitness prediction. It is intended as a research / prototyping toolkit for optimizing telescope pointings against probabilistic sky maps (HEALPix) and operational constraints.

🧭 Overview

• Purpose: evaluate and evolve observation schedules (chromosomes) composed of ordered pointings (genes). Fitness accounts for detection probability from HEALPix maps and operational costs (airmass, slew, overlap, visibility).
• Inputs: HEALPix probability map(s) and optional tile grids in data/.
• Outputs: console logs, Matplotlib / healpy visualizations (fitness evolution, best chromosome overlay), and a summary table of chromosomes and scores.

🧩 Main components

• 🧬 Chromosome representation

• Each chromosome is a list of genes. A gene is a dict containing id, ra_deg, dec_deg, start_time, exposure_s, and optional quality / probability fields.

• 📐 Derived fields

• recompute_derived_fields() computes per-gene airmass and slew (angular separation) using Astropy and stores them as airmass and slew_from_prev_deg.

• ⚖️ Fitness function (compute_fitness)

• Reward: HEALPix probability integrated over a small circular patch around each pointing (or per-gene probability if present), time-weighted to prefer earlier observations.

• Penalties: airmass above a threshold, slew-time (proportional to angular movement), overlap duration between pointings, and hard penalties for infeasible (below-horizon) pointings.

• Parameters: AIRMAX, LAMBDA_SLEW, LAMBDA_AIRMASS, LAMBDA_OVERLAP, LAMBDA_HARD located in the notebook for easy tuning.

• ♻️ Genetic Algorithm operators

• mutate() — small random perturbations to RA/Dec or start time.

• crossover() — single crossover point mixing parents.

• evolve_population() — selection (retain fraction), elitism, breeding, and mutation.

• 🤖 Neural-network surrogate

• Encodes chromosomes into fixed-length feature vectors (encode_chromosome) and trains a PyTorch MLP (FitnessNN) to regress real fitness values.

• After training, the NN predicts fitness for chromosomes and an NN-assisted GA loop demonstrates evolution using predicted fitness.

▶️ How to run

1. Ensure required Python packages are installed: numpy, astropy, healpy, pandas, matplotlib, torch, python-dateutil.
2. Open and run testing.ipynb in Jupyter / VS Code. The notebook is organized to run end-to-end, but you can selectively execute sections (data load, GA-only loop, NN training, NN-assisted GA).

Example environment setup (pip):

python -m pip install -r requirements.txt
# or install packages individually:
pip install numpy astropy healpy pandas matplotlib torch python-dateutil

📁 Files referenced by the notebook

• data/LALInference_v2.fits.gz (HEALPix probability map used by default in the example)
• data/GW170817_Generic_GRID.txt (optional tile grid used by make_population)

🔎 Key functions to review

• normalize_chromosome() — normalizes input gene fields and datetimes.
• recompute_derived_fields() — computes airmass and slew from prior pointing.
• compute_fitness() — core reward / penalty aggregation used by the GA.
• evolve_population(), mutate(), crossover() — GA operators.
• encode_chromosome(), FitnessNN — NN encoding and model.