Nishtha
Gupta

Most urban simulations script how developers behave. The city they produce then mostly restates the modeller’s own assumptions. This paper proposes a multi-agent reinforcement learning framework in which developer agents learn their strategy instead, so urban form emerges from an encoded economy of land value, rent and cost.

The trained agents also work as a diagnostic: what they learn to do shows which assumptions a simulation has built in. The paper extends the bottom-up urban simulation work I began at the AA.

Can coastal cities grow with rising water instead of only defending against it? Taking Amsterdam’s incremental canal development as a precedent, the thesis turned that question into a rule-based framework for climate-adaptive coastal growth.

I led the 4D land-use simulation and the urban research behind its rules. At its centre is a multiplayer city game: a communal player builds infrastructure from taxes, while developer players buy parcels, build and collect rent, and land values recalculate after every move. Building types change with distance from the water, from stilts and raised roads at the tidal edge to ordinary construction inland.

I also ran water-flow simulations across typologies, canal lengths and terrain, then built a physical water tank to test canal geometries under changing currents and heavy rain.

I connected Claude to Rhino, Grasshopper and ComfyUI and am taking a single hypothetical project from site to render with it. Each part tests one stage of the design process, on a real plot in Haryana, India.