Closed MadcowD closed 4 years ago
@MadcowD here are some quick "this vs that" comparisons
Garage | rllib |
---|---|
Focuses on single-machine environments for now, but allows you to add distributed components like sampling | Focuses aggressively on cluster environments, requiring you design your algorithm for clusters from the start |
Community-supported | Supported by Berkeley RISE lab and a new start-up called Anyscale |
Stand-alone project | Technology demonstration project for the larger ray distributed computing library |
No A2C family algorithms yet | Has A2C family algorithms |
Fewer, simpler abstractions | More, more complex abstractions (builders, mixins, hooks, trainers, etc.) |
No built-in hyperparam tuning interface | Built-in support for "ray tune" tool |
Generally easier to override and modify | Generally harder to override and modify (according to reports I've received) |
Supports both PyTorch and TensorFlow | Supports both PyTorch and TensorFlow |
No (explicit) multi-agent support yet | Multi-agent support |
Multi-task and meta-RL support | No multi-task and meta-RL support |
Thank you!!!!!
@ryanjulian can I ask what the barriers to A2C style algorithms are?
No technical barriers, just time. We spend a lot of time focusing on RL algorithms for continuous control, and A2C-style algorithms are more useful for gameplay domains. They would not be difficult to implement in garage, especially because we already have pixel-DQN.
Is it possible to train Garage DQN algorithm with multiple nodes using CPUs only? If yes, can you share a demo script or program. My sampling works at 2 nodes, 56 PPN using Ray Sampler.
Hi there,
I was wondering if anyone had established a reasonable comparison to rllib (ray). I am trying to chose between this and that and am having trouble making that decision.