A 8.81 TFLOPS/W Deep-Reinforcement-Learning Accelerator With Delta-Based Weight Sharing and Block-Mantissa Reconfigurable PE Array.
Sanghyuk AnJunha RyuGwangtae ParkHoi-Jun YooPublished in: IEEE Trans. Circuits Syst. II Express Briefs (2024)
Keyphrases
- reinforcement learning
- field programmable gate array
- programmable logic
- systolic array
- low cost
- reconfigurable architecture
- information sharing
- function approximation
- compute intensive
- state space
- markov decision processes
- hardware implementation
- parallel implementation
- learning process
- knowledge sharing
- machine learning
- reinforcement learning algorithms
- multi agent reinforcement learning
- learning algorithm
- temporal difference
- digital images
- markov decision process
- share information
- dynamic programming
- digital signal
- robotic control
- image processing
- multi objective evolutionary