Wrong suming output in simulink r2023A

Greetings!

We are training a reinforcement learning agent using the R2023a version in Ubuntu 22.04 LTS. We have built a reward block that should result in a reward value of -150 when the input is under a particular range. We have observed that the input to the summing block in the reward function is of value -150 (-100, -50, 0, 0) while the output of the sum block is of value 106

When we tested with constant values, as shown in the attached image, the sum block resulted in a value of -150, as intended.

This has affected the training of agents drastically.

Answer 1

[Steven Lord]

I don't work with Simulink that much, but I'd check the data types of the various signals in your code.

106 - (-150)

ans = 256

The fact that these two numbers differ by a power of 2 makes me suspicious that there's some conversion with wrapping going on.