Currently, the step-size change in CSA, $\exp\left(\frac{c}{d}(||p_\sigma|| / \chi_n- 1)\right)$, is a factor clipped into $1/e$ and $e$ via the hard coded attribute .max_delta_log_sigma = 1. This is however undocumented!
The option
{'CSA_clip_length_value': 'None #v poorly tested, [0, 0] means const length N**0.5, [-1, 1] allows a variation of +- N/(N+2), etc.'}is related and restricting the size of a single summand in $p_\sigma$.
Arguably, step-size clipping is universal and could or should be handled independently of the control method?
Currently, the step-size change in CSA,$\exp\left(\frac{c}{d}(||p_\sigma|| / \chi_n- 1)\right)$ , is a factor clipped into $1/e$ and $e$ via the hard coded attribute
.max_delta_log_sigma = 1. This is however undocumented!The option
{'CSA_clip_length_value': 'None #v poorly tested, [0, 0] means const length N**0.5, [-1, 1] allows a variation of +- N/(N+2), etc.'}is related and restricting the size of a single summand in$p_\sigma$ .
Arguably, step-size clipping is universal and could or should be handled independently of the control method?