Making Things Up: The Power and Peril of PCG (GDC)

• Speakers: Gillian Smith (Northeastern University) and Julian Togelius (New York University), names transcribed cleanly. The title is from the video listing; the transcript does not say it. A third speaker ("Nathan") follows them; his part is not in this transcript.
• Year: not stated. 2015 is implied: the PCG Jam ran the previous November and returns in 2015, and an AI game jam is due at the end of March. It appears to be a GDC AI Summit session (Tanagra was shown at an earlier AI Summit).
• Video: https://www.youtube.com/watch?v=B11RlHZsmGE
• Doubtful spellings in the transcript, reconstructed: "Paul Tussaud's City Conquest" (Paul Tozour), "Cameron Brown's Ludi" and "Yabaloth" (Cameron Browne, Yavalath), "Joris Dorman" (Joris Dormans), "rope bossom" (Ropossum, for Cut the Rope).

The talk's lessons, in short

1. Four families of generators, each with a trade-off. Constructive (Spelunky's hand-made tiles plus a decoration pass): cheap and tailored, but players spot its patterns quickly. Grammars: much from little, but they overgenerate regular structures and need a filter that throws bad results away. Constraint solvers (Refraction, Tanagra): guarantees such as playability, but every common-sense rule must be written down, and a conflicting set fails without saying why. Search or evolution: general, slow, and only as good as the fitness function; defining it as fun is a trap.
2. Combining families often works best (an evolved grammar for Mario levels, constraints plus a grammar in Tanagra).
3. Choose by building blocks (authored chunks versus small components), speed (online or offline), and the control players and designers get: parameters (a Civilization map setup), indirect control (Galactic Arms Race evolving the weapon you use most), experiential control (Tanagra's pacing timeline) or compositional control (50% ocean).
4. Mixed-initiative tools let a designer and a generator take turns: Tanagra, Sentient Sketchbook (suggestions and balance meters for strategy maps), Ropossum (generates and solves Cut the Rope levels).
5. Pressing generate a few times is not a sample of the generator's space: a hundred coin flips can come up 90 heads.
6. Expressive range: define numerical metrics, generate a lot (overnight for slow generators), "score each piece of content that you create and map out with a 2D histogram that can be visualized as a heat map". Their platformer metrics: leniency (each piece weighted by the harm of failing it; a gap kills, a Goomba costs a power-up), linearity, challenge density, and pattern density and variation (closeness to the original Mario levels).
7. Maps of different generators show some reach regions others never can; maps of one generator under different parameters uncover biases nobody intended.

What the kit took, and where

• examples/csharp/src/Example.Core/Procgen/ExpressiveRange.cs: Metric<T> (name, score, fixed range, so two maps share axes) and ExpressiveRange.Map(generate, seeds, x, y, bins), a bins x bins count grid with OccupiedShare, PeakShare, PeakBin, MeanX/MeanY, Overlap and Unreached to compare two maps, Summary() for the log and ToCsv() (every bin, culture-independent) for a heat map. A score outside its range throws, naming the seed.
• examples/csharp/tests/Example.Core.Tests/ExpressiveRangeTests.cs: 7 tests. A toy platformer row (60 segments of ground, gap or enemy, stepping up and down) scored on the talk's leniency (gap 1, enemy 0.5) and a least-squares linearity, mapped for two parameter sets over seeds 1 to 2000; and the kit's Eldritch maze (Maze.cs) mapped on dead-end share against junction share for its newest-room knob at 0 and 100%.
• .agents/skills/procgen/SKILL.md: a "Map the expressive range" test rule.

Not taken

• The four families' trade-offs, mixed-initiative tools and the control taxonomy: design advice with nothing to test; the skill's technique table already picks by content. No constraint solver or evolutionary search is added.
• Challenge density and the two pattern metrics: they need a corpus of reference levels; the helper takes any metric.
• Plotting: the CSV is the artefact; any spreadsheet or plotting tool draws the heat map.

Measured (10 x 10 bins over 0..1)

• The two row configs share 51% of their levels by bin (Overlap); 15 bins the first reaches the second never does, while their means differ by under 0.05: the mean hides the bias, the map shows it.
• The bins of the first ten seeds hold 45% of the 2000 levels: ten presses never showed the kind of most levels made.
• The two maze knob settings share no bin (Overlap 0).
• Proof: the tests fail when Map ignores the y metric (4 of 7), when out-of-range scores are clamped instead of thrown, and when the CSV uses the current culture (de-DE writes decimal commas). All 7 run in about half a second.