Evolved scheduling policies beat handcrafted heuristics under uncertainty

Handcrafted scheduling heuristics fall apart when satellite task visibility, operational profits, and onboard resource consumption fluctuate dynamically.

Static Earth observation scheduling assumes fixed execution windows and deterministic parameters. In reality, actual operational conditions cannot be known beforehand. The Uncertain Agile Earth Observation Satellite Scheduling Problem (UAEOSSP) addresses this gap directly, framing scheduling around the inherent volatility of task profit, resource consumption, and task visibility.

Attempting to resolve this with hardcoded heuristics produces rigid dispatch behavior. To tackle the problem dynamically, an effective Genetic Programming Hyper-Heuristic (GPHH) automates the generation of scheduling policies. Instead of forcing operators to manually design heuristics for every scenario, GPHH evolves policies that satellite systems can utilize directly to adjust operational plans in real time.

The empirical performance demonstrates clear margins over conventional design approaches:

• Evolved scheduling policies achieve an average improvement of 5.03% compared to well-designed Look-Ahead Heuristics (LAHs).
• Compared to Manually Designed Heuristics (MDHs), the policies generated by GPHH achieve an average improvement of 8.14%.

Where it breaks: the evolved policy is strictly an artifact of the hyper-heuristic formulation and the uncertain factors it accounts for. If in-flight conditions exhibit dynamics beyond the modeled uncertainties of profit, visibility, and resource consumption, the policy's decision quality is untested. Furthermore, these performance margins are measured against LAHs and MDHs; deploying GPHH requires accepting the evolutionary computation overhead to generate the policies before gaining the runtime execution benefits.

The takeaway: do not waste engineering cycles handcrafting dispatcher heuristics for agile systems operating under uncertainty. Automate policy generation with hyper-heuristics, then execute the evolved scheduling policies in real time to adapt to unknown operational states.