Using Simulation to Train and Test Artificial Intelligence for Business Applications

Software architectures used for DRL in simulation models

Different software architectures used for deep reinforcement learning with simulation models. Learn more in the presentation.

The relationship between simulation and artificial intelligence is increasingly close, especially for deep reinforcement learning. The AnyLogic Company CEO, Dr. Andrei Borshchev, explores this trend in Using Simulation to Train and Test Artificial Intelligence for Business Applications, a presentation given at the GE EDGE & Controls Symposium 2019 in Niskayuna, NY.

Below is the video recording and a brief summary. You can also find the video on our YouTube channel with timestamped topic sections.

Using Simulation to Train and Test Artificial Intelligence for Business Applications.

The presentation begins with a quick look at the type of simulation AnyLogic provides and also dispels some misconceptions about digital twins (4m 46s). A point illustrated with decisionLab’s industrial digital twin for Siemens - ATOM (10m 37s) (see also, the case study).

Why AI and Simulation?

This question marks the start of the main theme (14m 8s) of the presentation and shows how simulation fits in with the different types of AI technologies, as well as the different ways simulation can be used with AI:

  • To generate synthetic training data
  • To provide learning environments
  • As a testbed for trained AI

For these, it is possible to see how simulation helps when data is hard to get, maybe because it was not collected, or collection is too dangerous or costly, or simply because a real source of data does not exist yet. The great benefits AI practitioners are finding in simulation come from its low-cost and risk-free nature.

The provision of learning environments for training reinforcement policies is highlighted as a rapidly developing area and the presentation provides both a technical (21m 21s) (Machine Learning vs Optimization for Traffic Lights) and a commercial example (29m 53s) (Industrial Problem Resolved by AI and Simulation).

To wrap up, the conclusion (32m 26s) highlights three key challenges facing integrated simulation and AI. These center around the difficulties with scaling, the skills required for effective simulation modeling, and problems with the development process.

The challenges are not insurmountable, however. There are ways forward – in particular, new tools, such as Pathmind, are beginning to ease development processes.

Take a look at the video and see how the worlds of simulation and artificial intelligence are coming together and changing. Leave a comment below: What are your experiences working in these two fields? Do you agree with the challenges above? Let us know!

Related posts