NativeRL enables you to apply Reinforcement Learning (RL) to AnyLogic or Python-based simulations. We have optimized NativeRL to support almost all industrial use cases in verticals such as supply chain, manufacturing and many more.
- Product Delivery - https://cloud.anylogic.com/model/23213598-1bc4-419a-ade2-3b0e8a220e0d
- Flexible Manufacturing - https://cloud.anylogic.com/model/9044fa72-c5dd-4079-9f6a-a877464b905f
- AGVs - https://cloud.anylogic.com/model/bc88f3be-b5a6-4962-abc1-a26def59994d
- Mine Pit Operations - https://cloud.anylogic.com/model/a5c49106-d286-4f2a-9723-83d4a996672b8
To get started, please navigate to our wiki. There are generally 3 steps to get started.
- Build the NativeRL libraries.
- Install NativeRL locally and add Pathmind Helper to your AnyLogic simulation.
- Train your Policy.
RL Algorithms
Currently, NativeRL only supports PPO. From our experience, PPO works successfully for all use cases so there has not been a need to experiment with other RL algorithms.
Hyperparameter Tuning
Both PBT and PB2 are supported out-of-box. From our experience, PBT more consistently converges as compared to PB2. As such, PBT is the default trial scheduler in NativeRL.
| Hyperparameter Tuning Algorithm | Background |
|---|---|
| Population-Based Training (PBT) | https://www.deepmind.com/blog/population-based-training-of-neural-networks |
| Population-Based Bandits (PB2) | https://www.anyscale.com/blog/population-based-bandits |
RL Implementation
All features below can be mixed and matched to fit your specific use case. For more information about multi-agent implementations, please see https://bair.berkeley.edu/blog/2018/12/12/rllib/.
| Feature | Expected Usage | Status |
|---|---|---|
| Single Agent / Single Policy | The simplest possible RL implementation. Often used in toy examples. | Supported and Validated |
| Multi-Agent / Single Policy | Crucial for multi-agent simulations (e.g. controlling 10 AGVs simultanously). Multi-agent RL is often necessary for real-life industrial use cases. |
Supported and Validated |
| Multi-Agent / Multi-Policy | We have not identified any real-life industrial use cases that require multiple RL policies simultanously. | Unsupported |
| Discrete Action Spaces | Discrete choices such as an on/off switch. | Supported and Validated |
| Continuous Action Spaces | Continuous ranges such as a steering wheel. | Supported and Validated |
| Tuple Action Spaces | Multiple simultaneous actions such as driving a car which needs to stop/go and turn. | Supported and Validated |
| Action Masking | Prevent RL from taking illegal actions. Crucial for helping RL learn in use cases which contain many illegal actions. | Supported and Validated |
| Auto-Regressive Actions | Supported but we haven't found a good use case to validate this feature. Auto-regressive actions are often used in use cases such as robotics arms. | Supported but Untested |
NativeRL enables using environments for Reinforcement Learning implemented in Java or C++ with Python frameworks such as Ray's RLlib. You can think of it as a bridge between Java and Python.
It defines an intermediary C++ interface via the classes in nativerl.h, which are mapped to Java using JavaCPP and the classes in the nativerl submodule. The nativerl::Environment interface is meant to be subclassed by users to implement new environments, such as the ones outputted from code generated by AnyLogicHelper.java. The C++ classes are then made available to Python via pybind11 as per the bindings defined in nativerl.cpp, which one can use afterwards to implement environments as part of Python APIs, for example, OpenAI Gym or RLlib, as exemplified in code generated by RLlibHelper.java.