FAIR introduced SaLinA, a framework for developing models of sequential decision-making. Possible applications include reinforcement learning, computer vision, natural language processing.
SaLinA’s idea is to make the implementation of sequential decision-making processes (including reinforcement learning methods) as simple as the implementation of neural network architectures. To do this, with the help of SaLinA, it is proposed to represent any task of sequential decision-making in the form of a chain of simple agents that sequentially process information.
Salina is an extension of PyTorch (however, its principles can be extended to other deep learning libraries such as JAX) and consists of only a few hundred lines.
The main advantages of SaLina:
- SaLina is easy to understand and use;
- SaLinA allows you to create complex agents by combining simpler agents with predefined containers;
- Salina provides an NRemoteAgent wrapper that can execute any agent in multiple processes, speeding up the calculations of specific agents. Used in addition to the algorithm running on the CPU or graphics processors of a computer with this library, it simplifies the scaling process.