Table 1:

Testing Values (Relative Regrets) Comparisons of ITRs

Value ╲ ITR	DR-ITR	Standard ITR	LB-ITR
Training ${\mathcal{V}}_1$				0.6253 (37.36%)	0.9982 (0%)	0.9982
Testing ${\mathcal{V}}_{1,\text{test}}$	4.8230 (9.16%)	0.2927 (94.49%)	5.3096

¹DR-ITR maximizes ${\mathcal{V}}_c^k(d)$ defined in (4) with $k=2$ and $c=20$ over the linear ITR class.

²Standard ITR maximizes ${\mathcal{V}}_1(d)$ over the linear ITR class.

³LB-ITR maximizes ${\mathcal{V}}_1(d)$ or ${\mathcal{V}}_{1,\text{test}}(d)$ over the linear ITR class.

⁴Values (larger the better) can be comparable within rows but incomparable between rows.

⁵Relative $\text{regret}(\text{ITR})=[\text{value}(\text{LB}-\text{ITR})-\text{value}(\text{ITR})]/|\text{value}(\text{LB}-\text{ITR})|$ (smaller the better)

⁶A size-10,000 sample is generated for fitting DR-ITR and LB-ITRs, and an independent size-100,000 sample is generated for evaluation under ${\mathcal{V}}_1$ and ${\mathcal{V}}_{1,\text{test}}$.