Towards Safe and Sustainable Reinforcement Learning for Real-Time Strategy Games

Information Sciences, Journal Year: 2024, Volume and Issue: 679, P. 120980 - 120980

Published: June 12, 2024

Combining Deep Neural Networks with Reinforcement Learning, known as Learning (DRL), is revolutionizing fields like medicine, industry, and gaming. DRL has achieved groundbreaking results, particularly in complex Real-Time Strategy (RTS) games such StarCraft II Dota 2, serving benchmarks for testing RL algorithms' robustness safety. Despite these successes, algorithms face challenges, including high computational costs a lack of safety-aware approaches. Training requires extensive resources, leading to significant divide between developed on supercomputers those feasible standard hardware. This also raises sustainability concerns due increased CO2 emissions. Additionally, most are risk-neutral, limiting their deployment safety-critical systems. We present novel model-based approach, the Safe Observations Rewards Actions Costs Ensemble (S-ORACLE), address challenges. S-ORACLE balances robust safety awareness minimized risk efficiency. Empirical validation across game environments—Deep RTS, ELF: MiniRTS, MicroRTS, Warehouse, II—demonstrates that outperforms state-of-the-art methods by significantly improving performance, reducing costs, lowering environmental impact, while maintaining efficiency adaptability training.

Language: Английский

Learning Memetic Algorithm Based on Variable Population and Neighborhood For Multi-Complex Target Scheduling of Large-Scale Imaging Satellites

Published: Jan. 1, 2024

Language: Английский

Learning memetic algorithm based on variable population and neighborhood for multi-complex target scheduling of large-scale imaging satellites

Swarm and Evolutionary Computation, Journal Year: 2024, Volume and Issue: unknown, P. 101789 - 101789

Published: Nov. 1, 2024

Language: Английский