An Artificial Intelligence Approach to Token-Efficient Memory Architecture for Deep Reinforcement Learning via Learned Summarization and DynamicMemory Mechanisms

Abstract

Tentu, ini dia versi satu paragraf yang dirapikan:

"Deep Reinforcement Learning (DRL) has achieved impressive success in sequential artificial intelligence decision-making problems, yet challenges remain in addressing long-term dependencies and computational efficiency. This paper explores token-efficient memory schemes to enhance DRL performance in high-dimensional, partially observable AI systems. We propose the Token-Efficient Memory Architecture (TEMA), a dynamic declarative-procedural memory switching architecture that combines learned summarization with this switching mechanism. TEMA introduces periodic summarization tokens, creating compact vector representations of historical trajectory segments, achieving an 8-fold compression of context with less than a 2% performance loss compared to full-context baselines. Benchmarking memory-augmented models like RNNs, Transformers, NTMs, and TEMA on robotic control (MuJoCo), real-time strategy (StarCraft II), and multi-agent settings (Pommerman) reveals TEMA to be 21% more effective than LSTM baselines while using 40% less memory on GPUs than conventional hybrid designs. Our synopsis system facilitates better credit allocation over longer time horizons, overcoming the key memory-performance trade-off that has limited the practical use of memory-enhanced DRL systems, positioning it as a leading AI method for DRL problems, memory architectures, token efficiency, learned summarization, long-term dependencies, and partial observability."