Skip to content

Appendix B: Formula Sheet

Expected value:

\[ E[X] = \sum_i p_i x_i \]

Translation:

probability-weighted average outcome

Variance:

\[ \mathrm{Var}(X)=E[(X-E[X])^2] \]

Translation:

spread around the expected value

Monte Carlo estimate:

\[ E[F(x,\omega)] \approx \frac{1}{N} \sum_{n=1}^N F(x,\omega^n) \]

Translation:

average the objective over sampled futures

Expected value-variance score:

\[ \text{score}(x)=E[F(x,\omega)]+q\,\mathrm{Var}(F(x,\omega)) \]

Constraint-violation penalty:

\[ \text{penalty}(x,\omega)=p\max\{\text{load}(x,\omega)-\text{capacity},0\} \]

Hurwicz criterion:

\[ H(x)=(1-\lambda)\cdot\text{worst}(x)+\lambda\cdot\text{best}(x) \]

Minmax regret:

\[ \min_x\max_s\left(F(x,s)-\min_{x'\in X}F(x',s)\right) \]

Bellman decision:

\[ x = \arg\min_{x \in X(S_k)} \left[ R(S_k,x) + V(S_k^x) \right] \]

Maximization variant:

\[ x = \arg\max_{x \in X(S_k)} \left[ R(S_k,x) + V(S_k^x) \right] \]

Post-decision expected value:

\[ V(S_k^x) = \sum_{S_{k+1}} P(S_{k+1} \mid S_k,x) V(S_{k+1}) \]

State transition:

\[ S_{k+1}=S^M(S_k,x,\omega_{k+1}) \]

Translation:

next state equals old state plus decision effect plus newly revealed information

RHO:

\[ x = \arg\min_{x \in X(S)} R(S,x) \]

Course interpretation:

standard RHO is Bellman with V = 0 unless a future approximation is added

CFA:

\[ x = \arg\min_{x \in X(S)} \left[ R(S,x) + I(S,x) \right] \]

Constraint-shaping variant:

\[ (A+\Theta)x \le b+\theta \]

VFA:

\[ x = \arg\min_{x \in X(S_k)} \left[ R(S_k,x) + \hat{V}(S_k^x) \right] \]

Feature / aggregation version:

\[ \hat V(S_k^x)\approx \hat V(A(S_k^x)) \]

where \(A(S_k^x)\) is a smaller feature vector such as \((\text{time},\ \text{remaining capacity})\).

VFA update:

\[ \hat{V}_{new} = \hat{V}_{old} + \alpha \left(\text{observed value} - \hat{V}_{old}\right) \]

PFA threshold:

\[ \pi_\theta(S_k)= \begin{cases} \text{accept}, & \text{if money/space}>\theta\\ \text{reject}, & \text{otherwise} \end{cases} \]

Look-ahead score:

\[ \text{score}(x)=R(S_k,x)+\frac{1}{N}\sum_{n=1}^{N}J^n(\text{future after }x) \]

Supervised learning:

\[ f(z) \approx y \]

Sample-policy evaluation:

\[ \widehat{J}(\pi) = \frac{1}{N} \sum_{n=1}^N J^n(\pi) \]