This is the final exam for Fitness Landscapes. It pulls the whole course together: Sewall Wright’s picture that turns improvement itself into movement across a terrain, where every possible design is a point, similar designs sit side by side as neighbours, and height means fitness. From that one map falls the devastating fact the whole course orbits — evolution and every trial-and-error search can only walk uphill, so they climb the nearest hill and stop, marooned on a local optimum below a global optimum they cannot reach because the only route there runs downhill first. You’ll be tested on the vocabulary of peaks, valleys, ridges and neighbours; why hill-climbing is greedy, local and foresight-free, and why that one rule is both its power and its prison; the biological fingerprints of a trapped climber — the vertebrate eye’s backwards retina and blind spot, the giraffe’s recurrent laryngeal nerve, the panda’s thumb — and the lock-in of QWERTY; what makes a landscape smooth or rugged (interactions between parts, i.e. epistasis) and the forces that cross valleys — mutation and big jumps, genetic drift in small populations, recombination, and neutral ridges; the Red Queen twist where the terrain itself won’t hold still; and the transfers out of biology — gradient descent, business disruption, skill plateaus, and the master trade-off of all search, explore vs. exploit. Take your time; several questions look easy until you spot the trap — “it improved, so it must have found the best,” “these flaws prove bad design,” “just keep optimizing what already works.”
How this exam works
Read carefully — this exam is final. Each question appears one at a time. Once you submit an answer it is locked for good: there’s no going back, no retry, and no restart. Your score is hidden until the end, where you’ll see a pass/fail verdict. The pass mark is 70%. A few questions ask you to select all correct answers.
What exactly is a fitness landscape?
Select an answer to continue.
Course Recap
Big picture
Fitness landscapes, in one picture
- Fitness Landscapes
- The map & its vocabulary
- Spread every possible design on a map so similar designs are neighbours, then make height = fitness: you get a terrain of peaks (better than all neighbours), valleys (low ground between peaks), ridges (high connecting paths), one global optimum (tallest peak) and many local optima — and the axes are design space, not time, while height is fitness, not complexity
- Hill-climbing & its prison
- Evolution and every trial-and-error search keep helpful changes and drop harmful ones, so they only step uphill — greedy, local, and foresight-free — which reliably climbs the current hill (its power) but bars them from crossing any valley to a taller hill (its prison), because the route there runs downhill first
- Local optima & their fingerprints
- A search settles on the first reachable peak and stays because every neighbouring step is worse, so "good enough" beats "best" — leaving tell-tale kludges: the backwards-wired retina and blind spot (vs the octopus eye), the giraffe's recurrent laryngeal nerve detour, the panda's thumb, and the QWERTY lock-in of path dependence
- Ruggedness & crossing valleys
- Interactions between parts (epistasis) fracture a smooth hill into a rugged range of traps; the valley-crossers are mutation/big jumps (most land in a valley — that's the cost), genetic drift in small populations, recombination shuffling good pieces, and neutral ridges that let a search wander sideways for free until a taller hill comes into reach
- The Red Queen & the model everywhere
- The Red Queen makes the terrain itself shift — a conquered peak subsides as rivals adapt, so there is no final summit; and the same map governs gradient descent (local minima, learning-rate step size, simulated annealing, random restarts), business disruption (a rival on a taller mountain), skill plateaus (a step down to climb higher), and the master trade-off of all search — explore vs. exploit
- The map & its vocabulary
Key takeaways
A fitness landscape is Sewall Wright’s map of every possible design — neighbours laid side by side, height = fitness (not time, not complexity) — and from it falls one relentless fact: evolution and every trial-and-error search can only walk uphill. Because hill-climbing is greedy, local and foresight-free, it climbs whatever hill it starts on and stops at the top, marooned on a local optimum below the global optimum, unable to cross a valley because the route there runs downhill first — and “downhill” means worse, and a hill-climber won’t take a worse step. That single rule is both its power (relentless improvement on the current hill) and its prison, and it leaves fingerprints all over biology: the backwards-wired retina and blind spot, the giraffe’s absurd recurrent laryngeal nerve, the panda’s thumb, the QWERTY lock-in — each a workable-but-flawed local peak, not a mistake but a signature of the process. Landscapes range from smooth (one gentle hill, easy) to rugged (a maze of traps born from epistasis, the interaction between a design’s parts), and the only ways across a valley are the valley-crossers — mutation and big jumps (most land in a valley, which is their cost), genetic drift in small populations, recombination, and neutral ridges — never plain uphill climbing. The Red Queen then refuses to let the terrain sit still: conquer a peak and it subsides as rivals adapt, so there’s no final summit. And the map is ruthlessly general — it’s gradient descent in machine learning (local minima, learning-rate step size, simulated annealing, random restarts), disruption in business (a rival on a taller mountain you can’t climb to), the skill plateau that demands a step down to go higher, and the master trade-off of all search: explore vs. exploit — climb the hill you’re on, or jump to look for a better one.