Module 11 · Exam Readiness

Oral Exam Bootcamp

Everything converges here: both papers side by side, the themes examiners will probe across them, a technique for answering under pressure, a 25-question bank to rehearse with, and a final mock exam spanning the whole course.

▶

Audio recap

A ~2-minute spoken summary of this module — great for revision on the go.

11.1 The two papers side by side

An examiner can pivot between papers in one breath. This table is the map — know every cell cold.

	Paper 1 — du Plessis, Strauss & Rix (2021)	Paper 2 — Hu, Lin & Lipson (2025)
Problem	Can macro-level forecast models capture the low-level power dynamics of a large PV plant?	Can a robot learn morphology, kinematics & motor control from a single 2D camera watching itself?
Domain	Utility-scale solar: 75 MW plant, 84 inverters, Northern Cape (South Africa)	Robotics: 4-DOF arms (two morphologies + one damaged variant), sim (PyBullet) + real
Task type	Supervised multi-target regression: 21 × 15-min steps, 1–6 h ahead	Self-supervised regression: per-query density σ and visibility α, trained via rendered silhouettes
Input data	On-site sensor time series: power, GHI, temperature (+ engineered solar angles β/Φ, one-hot wind-dir/month); no weather forecasts	Motor-babbling video frames (100×100, colour-segmented) + joint angles from encoders
Architecture(s)	FFNN, LSTM-RNN, GRU-RNN (sliding windows 1 h / 3 h / 6 h)	FFKSM: three fully connected networks (coordinates encoder, kinematic encoder, predictive module), NeRF-inspired, with rotation pre-transform + positional encoding
Loss	MSE on power (NRMSE for selection; MAE penalizes large errors less, so rejected)	MSE between rendered and segmented binary silhouettes, (1/WH)ΣΣ(Pred−GT)²
Optimizer	Adam, lr 1×10⁻⁴, ReLU, early stopping (patience 20), no dropout benefit	Adam + ReLU for training; Adam again (lr 0.04) for gradient-based control
Data scale & split	4 years, chronological 2/1/1 (train/val/test); cleaned by elimination, interpolation, imputation	12,000 frames: 10,000 train/val at 8:2 + 2,000 held-out test
Key methodological invention	Inverter clustering (Euclidean distance + K-means, K=10) with representative-inverter optimisation; 3-phase unbiased hyperparameter framework	Learned visibility output (vs NeRF's transmittance) + motion planning by gradient descent through the differentiable self-model
Evaluation	NRMSE / MAE / MAPE by weather type + bootstrap 95% CIs (m = 10,000)	2D image MSE (px², n = 2,000) vs RS and NN baselines; qualitative sim + real overlays
Headline result	Best macro GRU 8.12% NRMSE; inverter-cluster GRU 8.02% — improvement real but marginal, biggest for ≤3 h and intermittent weather	0.004 px² whole-body error (2.5× better than NN); spiral tracking & collision-free planning without kinematic equations; damage recovery by fine-tuning
Main limitation	Single flat, arid site; generality to uneven or larger plants unproven	2D-image-only quantitative evaluation (no direct 3D ground truth); controlled visual environment

Don't mix the two Adams of Paper 2 Training the FFKSM uses Adam on network weights; motion planning uses Adam at lr 0.04 on joint angles through the frozen model. Paper 1 uses Adam at lr 1×10⁻⁴ on weights only. Confusing these under pressure is a classic self-inflicted wound.

11.2 Shared themes examiners will probe

Cross-paper questions are where strong candidates separate from adequate ones. Six themes recur:

Same optimization core. Both are regression with MSE loss, Adam, ReLU activations, and a guard against overtraining (explicit early stopping with patience 20 in Paper 1; early aborting/restarts and staged training in Paper 2). Say this early in any comparison answer — it frames everything else as variations on one machine.
Clever input engineering — the SAME lesson twice. Paper 1 injects known physics as features: future solar altitude β and azimuth Φ (deterministic astronomy — why force a network to learn the sun's path?) and one-hot wind-direction/month (raw degrees lie: 359° ≈ 0°). Paper 2 injects known geometry as structure: the first two joints' rotations are applied analytically (X′ = T⁻¹X) and coordinates are positionally encoded (3→33, 2→22). One principle: inject the structure you already know, so the network spends its capacity only on the unknown.
Both fight a degenerate shortcut. Paper 1's worry: a macro model could score well while being blind to low-level dynamics — so they built 84 inverter-level models to check. Paper 2's failure mode: predicting all-black images — fixed by restart-on-stagnation and focusing early loss on the central 50×50 region. In both cases the cure was restructuring the training/evaluation signal, not a bigger network.
Both quantify honesty. Paper 1: bootstrap 95% CIs (m = 10,000), full-year test set across all weather types, and an explicit taxonomy of why other papers' superiority claims mislead. Paper 2: a genuinely strong NN baseline, a 2,000-sample held-out test set, and an admitted evaluation compromise. Neither paper claims more than its evidence.
Both reuse similarity. Paper 1 clusters inverters by Euclidean distance on their power histories so 10 representative models stand in for 84. Paper 2's NN baseline retrieves by joint-angle proximity — and its whole premise is that nearby configurations look similar. Similarity as a computational shortcut vs similarity as a baseline: same intuition, opposite roles.
Both end in a practical verdict. Paper 1: macro models are sufficient — practitioners with <10 MW plants can skip low-level modelling. Paper 2: self-models make damage recovery autonomous. Applied papers earn their place with actionable conclusions; say them unprompted.

11.3 How to answer in an oral exam

Structure beats brilliance under pressure. For every question:

1 · Direct answer (one sentence)→ 2 · Mechanism (why)→ 3 · Evidence with a number→ 4 · Honest limits

Two rules of survival: never bluff a number — "I don't recall the exact figure, but the ordering was GRU best, then FFNN, then LSTM" is respectable, an invented decimal is fatal; and reason aloud when unsure — examiners grade your process, and silence gives them nothing to grade.

Worked example: "Why did the GRU outperform the LSTM here?"

Strong answer: "Directly: the GRU's streamlined gating helped on this data volume. Mechanism: a GRU merges the LSTM's forget/input gates and drops the separate cell state, so it has fewer parameters per unit — an easier optimisation problem when data, though four years long, is one site's worth. Evidence: the best GRU reached 8.12% NRMSE versus the LSTM's 8.23%, and notably the selected GRU used only 2 hidden layers against the LSTM's 3 — a simpler model won. Limits: the margin is small, this is one plant, and the FFNN at 8.19% sat between them — so I'd claim 'GRUs were marginally better here', not 'GRUs are better'."

Weak answer: "GRUs are just better than LSTMs — they're newer and faster." — no mechanism, no number, an over-generalization from one site, and 'newer' is not an argument. This is the difference the examiner is listening for.

11.4 The question bank

No answers here on purpose: rehearse aloud, then check yourself against the referenced modules. Aim to open every answer with one direct sentence.

Block A — Warm-up definitions (revisit Modules 1–2 and 7)

Define NRMSE. Why normalise by rated capacity rather than by mean power?
What makes Paper 2's training self-supervised rather than supervised in the usual sense?
What is an implicit (query-based) model of a body, versus an explicit mesh or CAD model?
What is a bootstrap confidence interval, and what does "m = 10,000 resamples with replacement" mean operationally?
What is a hidden unit, and why were hidden-unit counts chosen as powers of two in Paper 1?

Block B — Mechanism (revisit Modules 3, 5, 6 and 8)

Walk through the gates of an LSTM cell, then say precisely what a GRU merges or removes.
Explain NeRF volume rendering with transmittance — then explain why Paper 2 could not use it and what Pred_ij = Σ_kσ_ijkα_ijk does instead.
Describe Phase-2 "guided grid search" (coordinated descent): what triggers extending the search domain, and when does it stop?
How does K-means assign 84 inverters to 10 clusters, and what role does the representative inverter play afterwards?
Trace one full iteration of gradient-based inverse kinematics through the FFKSM, from joint angles to updated joint angles.

Block C — Design defense (revisit Modules 9–10)

Defend MSE over MAE as Paper 1's cost function. When would MAE be the better choice?
Why does Paper 1 deliberately exclude external weather forecasts from its inputs?
Defend K = 10: what do the Elbow method, CH-index and Gap statistic each contribute?
Why must the FFKSM output visibility, not just density? What experiment justifies it?
Why positional-encode coordinates (3→33) instead of feeding raw (x, y, z)?
Defend the chronological 2/1/1 split, and the 8:2 + held-out-2,000 split — same principle or different?

Block D — Critique (revisit Modules 9–10, limitation sections)

Paper 1 studied one flat, arid, 75 MW site. Which conclusions travel, and which might not?
Attack Paper 2's 2D-only evaluation: construct a failure case that pixel MSE would miss.
Paper 2's segmentation needs a painted robot on a clean background. How fragile is the whole pipeline to that assumption?
Paper 1's gain from inverter-level modelling was marginal (8.02% vs 8.12%). Is publishing a near-null result a weakness or a strength? Argue both sides.
Interpret Paper 1's CI widths growing from ~59.5 to ~74.8 kW with horizon. What does that mean for a grid operator at 6 h?

Block E — Synthesis (whole course)

Could Paper 1's plant use a NeRF-style spatial model of the field — irradiance as a queryable function over (x, y, t)? Sketch inputs, outputs, training signal, and what could go wrong.
Could Paper 2's robot use bootstrap confidence intervals? On what quantity, resampling what?
Both papers avoid claiming too much. Check Paper 2's evaluation against Paper 1's four false-superiority factors (metrics, weak benchmarks, incomplete test sets, biased optimisation effort): which does it satisfy, and where is it most vulnerable?
"Inject known structure so networks learn only the unknown." Give both papers' instances of this principle, then one instance from any other field you know.

11.5 Final advice

Night-before checklist

Retake every module quiz to 80%+. Wrong answers the night before are gifts — each one is a question you now won't miss live.
Rehearse both elevator defenses aloud (Module 9 §9.8 and Module 10 §10.8), timed to 60 seconds each. Aloud, not in your head — the mouth needs the rehearsal, not the eyes.
Know every number in the §11.1 table. If you can reproduce that table from memory, you cannot be ambushed on facts.
Prepare one critical question per paper — examiners often ask what you would probe. Suggested: for Paper 1, generality beyond one flat arid site (would inverter-level modelling matter more on the rolling-hills Les Mées-style farm the paper itself cites?); for Paper 2, segmentation fragility outside the lab (unpainted robot, cluttered background, changing light).
Sleep. A rested "let me reason through that" beats an exhausted recital every time.