Train AI agents with reinforcement learning

An integral part of any reinforcement learning setup is providing RL agents with a reliable simulated environment. This is best accomplished by using a powerful, general-purpose simulation software with fast, consistent, and streamlined connections to RL algorithms. For experts or researchers who want to use AnyLogic models as their training environment for reinforcement learning, there are three available services: AnyLogic Cloud’s interactive API, third-party platforms for automated RL, and the ALPyne library.

Use Cases Workflows & Tools

Case 1: Optimal control of complex dynamic systems

An integral part of any reinforcement learning setup is providing the AI agents with a reliable simulated environment. This is best accomplished using a powerful general-purpose simulation software with fast, consistent, and streamlined connections to RL algorithms. The policies learned from trainings can eventually be deployed the real system which the simulation model was built from.


Case 2: Verify and validate simulation model

At its core, the reinforcement learning training process is comprised of an artificial explorer that examines and scrutinizes all corners of a simulation environment. With an appropriate reward schema, this mechanism could be used to partially automate some commonly repetitive aspects of the verification and validation process, allowing more thorough testing of the robustness and fidelity of the simulation model. Although this approach is still at its infancy, it has the potential to become an integral part of the verification and validation process for all types of models.


Case 3: Comparing the efficacy and performance of different RL algorithms

There are repositories of standardized RL environments for researchers to test and compare their algorithms on comparable playing fields. However, these widely used environments do not provide the variety and complexity that are commonplace in real simulated systems. A general-purpose simulation platform can provide sophisticated training environments that are able to be easily customizable, yet also can provide varying levels of complexities and complications which are unique to each industry and applied scenario.


Case 4: Serving as a comparison metric to assess the efficacy of human-designed policies

Analysts can choose, design, or curate all sort of rule-based, algorithmic, or heuristic-based solutions. Having access to a baseline solution, in the form of an RL policy, is extremely valuable to shed light on the efficacy of curated and manually shaped solutions - especially when these solutions are for scenarios that an absolute optimum is unattainable.

Workflows and Tools

For experts or researchers who want to use AnyLogic models as their training environment for reinforcement learning, there are three available services: AnyLogic Cloud’s interactive API, third-party platforms for automated RL, and the ALPyne library.

All three of these options use the RLExperiment as their main connection to the simulation model, which means the model can easily be ported to any of the other options. This gives you the choice of starting with the most convenient or relevant option to your specific use case. As your project develops, you can easily migrate to a different, more suitable workflow.

AnyLogic Cloud and its interactive API

AnyLogic Cloud and its interactive API

Upload the simulation model to the AnyLogic Cloud and use the interactive cloud API to communicate with user-assigned AI frameworks.

This option is for experts with manually defined RL training code who wish to train using simulation environments hosted on AnyLogic Cloud. Owners of AnyLogic Private Cloud have access to an interactive Python API which takes care of running the models on a scalable, server-based platform.

Automated RL Environment

Automated RL Environment

Upload (or connect) the simulation model to an automated RL training and development environment.

AnyLogic actively collaborates with pioneers in automated RL - Microsoft Project Bonsai & Pathmind - to streamline the process of establishing and executing RL trainings in applied applications. These automated platforms allow end-users to abstract away from the tedious process of finding the best RL algorithms and training parameters. This allows subject matter experts, who are looking for practical and scalable platforms, to benefit from reinforcement learning without needing a computer science degree.

Connection with ALPyne

Connection with ALPyne

Connect exported AnyLogic models and communicate with AI frameworks in a local Python environment via ALPyne.

For those interested to test how a manually curated RL setup works with an AnyLogic model on a local machine, ALPyne provides a way do so. This Python-based package allows you to communicate with an AnyLogic model exported from the RL Experiment. ALPyne follows a similar API to AnyLogic Cloud’s interactive API, providing the opportunity to move to a more scalable setup with minimal refactoring.