Revisit PES error magnitude expectations

[tbekolay]

At https://github.com/nengo/nengo-loihi/blob/master/nengo_loihi/simulator.py#L498 it seems as though we are getting the error from the host, doing x = int(100 * x), and then sending it to the chip. There's a comment in that block of code saying >128 is an issue on chip; @drasmuss noted:

The comment says 128 but the code uses 100? Also is this assuming that errors will be <= 1, or is that handled elsewhere somehow?

If we are making an assumption about error magnitudes, we should be explicit about it. In my experience, error values are very rarely going to be this high, so we might get better performance by looking at a tighter range of error values (say from -0.2 to 0.2).

[drasmuss]

#139 added a configurable pes_error_scale parameter which can be used to adjust this. Do you think that's enough to consider this resolved?

[hunse]

The only open question for me would be default values. If error does in fact typically go between -0.2 and 0.2, then we'd want a larger default for pes_error_scale. But I think this is fairly problem dependent. For example, we had to scale it down for the adaptive control example, since that used larger errors: 32f8bd6. I think that 100 is a good default, since if the range of errors is smaller, then it'll still work (just with slightly worse resolution).

[drasmuss]

Yeah, assuming that errors are +/-1, with the .28 fudge factor added, seems like an unsurprising default at least, even if it may not be optimal.