From a tension to attributes and design variables¶
A tension is read in the conditioned metrics. Two routes lead out of them, and only one is a calculation.
To the attributes' own units¶
The conditioning is log-affine, log m = alpha log(a + s) + b, so a tension in the metrics
maps to the attributes by the chain rule: d s_j / d s_i = -(w_i alpha_i)/(w_j alpha_j) (a_j + s_j)/(a_i + s_i).
Give design_tensions the fitted preconditioner and the raw attributes and it returns this as
attribute_tensions.
import numpy as np
from tradespace import (
HyperbolicTradeSpace,
LogAffinePreconditioner,
design_tensions,
)
rng = np.random.default_rng(0)
S = rng.uniform(1.0, 5.0, size=(300, 3))
pre = LogAffinePreconditioner(orientation=[-1, -1, 1]).fit(S)
M = pre.transform(S)
est = HyperbolicTradeSpace(temperature=10.0).fit(M)
dt = design_tensions(est, M, preconditioner=pre, attributes=S)
print(dt.attribute_tensions[0].round(2))
The sign in attribute units is minus the sign of alpha_i alpha_j: negative for two attributes
of one orientation (two costs), positive for one of each. "Negative where both contend" is a
statement about the metrics only.
To the design variables¶
The fit never sees the configuration, and the evaluation behind the attributes is a black box. A tension is a derivative along the fitted reference surface, not the effect of changing a design variable. What you can compute is each sampled design's tensions beside its configuration, and a regressor from configuration to metrics, whose slopes at the sampled designs are estimates of the evaluation's own. Past the sampled designs, a tension translated into a design variable is a conjecture: propose the configuration, build it and measure it.