question

We are currently working to improve our Bonsai examples. Currently, the samples are mainly a demonstration of connecting Bonsai and FlexSim. The example shows a correct setup, but not a good setup.

In the case of the ChangeoverTimes example, there are currently several issues:

• The Reach goal is interfering with the learning process. A Reach goal tells Bonsai to get to the value in as few actions as possible. In this particular model, that directly competes with maximizing throughput; increasing throughput means increasing the number of actions to reach the goal. When I
• The Throughput observation is currently cumulative, but it needs to be reset between actions. With a cumulative value, Bonsai assumes that its actions get better and better later in the run, no matter what actions it takes.
• I suspect that the Maximize goal should have a higher value than zero
• The model itself isn't a very good learning environment. Once there are many items in the queue, Bonsai can always choose an item that matches the one previous, so that there is never a setup time again.

These issues are likely the cause of the behavior you are seeing.

As far as performance for unmanaged mode training, 5-6 iterations per second is expected. Unmanaged mode is meant for testing and debugging. For a full training run, create a Managed simulator. If an unmanaged simulator runs at 5-6 iterations per second, then a managed simulator will run at 15-20 iterations per second. In addition, Bonsai can use many instances of a managed simulator simultaneously.