Theory-grounded multi-agent recommendation

From an urban challenge,
to a defensible computational strategy.

Five cooperating agents parse an unstructured planning problem, retrieve relevant urban theory, resolve theoretical requirements against a corpus of computational methods, attach evidence-linked data sources, and stress-test the assembled strategy — entirely in your browser. No server, no waiting.

40urban theories
1898 – 2013
1,297algorithms
4 method families
1,381data sources
literature-mined
5agents
one pipeline
Run the pipeline ↓

01 — Live pipeline

Describe a challenge. Watch the agents reason.

loading knowledge base…

02 — Knowledge graph

Every recommendation is a subgraph, not a list.

Theories connect to algorithms through resolved computational requirements; algorithms connect to data sources through co-occurrence links mined from the literature corpus. Run the pipeline above, then drag, zoom and hover to interrogate the evidence structure behind the recommendation.

theory algorithm data source requirement match literature co-occurrence

— run the pipeline to build the graph —

03 — Knowledge base

2,718 entities, fully inspectable.

Nothing is hidden behind an API. The complete knowledge base ships with this page — search it directly.

    — select an entry —

    04 — Method

    Deterministic, auditable scoring.

    This page is a faithful client-side implementation of the published pipeline, adapted to the full bibliometric knowledge base. Identical input yields identical output; every score decomposes into inspectable parts.

    A1 Scenario analysis

    Keyword-profile domain classification over five urban domains; pattern extraction of objectives, constraints and stakeholders; complexity as a weighted six-component score Σ wᵢ·cᵢ over scope, actors and domain difficulty.

    A2 Theory retrieval

    Scenario and theory texts embedded with a compact urban-domain lexicon; cosine similarity ranks all 40 theories, with additive boosts for category–domain fit and application context, plus curated synergy rules (e.g. CPTED ↔ Defensible Space).

    A3 Algorithm matching

    Each theory contributes snake_case computational requirements; a token-level resolver maps them onto the capability vectors measured across 1,297 methods. Score = mean matched capability − 0.1·cost, shaped by requirement coverage; greedy selection enforces capability and family diversity.

    A4 Data source selection

    Weighted composite of literature co-occurrence with the chosen algorithms (0.35), reliability (0.20), domain–type fit (0.15), freshness from update cadence and age (0.15) and access complexity (0.15), with a diversity pass across source types.

    A5 Integration validation

    Fifty seeded Monte-Carlo perturbations of composite performance yield a robustness rate; a compatibility audit lists uncovered requirements; confidence combines component coverage and robustness minus a complexity penalty.

    KB Knowledge base

    Theories curated from the planning canon (1898–2013). Algorithms and data sources mined from an urban-AI literature corpus, retaining measured capability profiles, computational cost, reliability and co-occurrence structure.