Deep Reinforcement Learning for Unpredictability-Induced Rewards to Handle Spacecraft Landing

This paper proposes a descent reward (DR) system to enhance agent's navigation and landing capabilities in the LunarLander-v2 environment provided by OpenAI. The DR system is evaluated with Heuristics Control, Linear Function Approximation (LFA), and Deep Q-Networks (DQN) algorithms. Results demonstrate that the DQN algorithm with DR achieves a 97% average success rate, improving 17% over the Heuristic baseline and 37% over the default reward DQN in terms of fuel consumption. Furthermore, the resilience of DQN, Double Deep Q-Networks (Double-DQN), and Dueling Deep Q-Networks (Dueling-DQN) with the proposed DR system is evaluated. Despite increased unpredictability, results show the DR system effectively augments agent learning and decision-making. The Double-DQN approach exhibits superior stability and reward achievement compared to DQN and Dueling-DQN.