---
title: Mini Amusement Parks (MAPs) Benchmark
url: https://www.emergentmind.com/topics/mini-amusement-parks-maps
type: topic
---

# Mini Amusement Parks (MAPs) Benchmark

Mini Amusement Parks (MAPs) is an amusement-park simulator designed to evaluate decision making in a domain that combines open-ended objectives, sparse experience, long-horizon stochastic planning, and spatial reasoning. In its 2025 formulation, MAPs is presented as a benchmark for modelling business decisions, with human baselines and a comparative evaluation of LLM-based agents; humans outperform the tested systems by 6.5x on easy mode and 9.8x on medium mode. A related antecedent is Micro RollerCoaster Tycoon, an earlier open-source simulator that used MAP-Elites to explore diverse park layouts and theoretical questions about complexification and resource constraints in open-ended gameplay design [2511.15830] [2105.04342].

## 1. Scope, domain, and research framing

MAPs is motivated by the claim that practical domains such as business management require optimizing an open-ended and multi-faceted objective, actively learning environment dynamics from sparse experience, planning over long horizons in stochastic settings, and reasoning over spatial information. The benchmark is intended to assess an agent’s ability to model its environment, anticipate long-term consequences under uncertainty, and strategically operate a complex business. The published codebase is available at `https://github.com/Skyfall-Research/MAPs` [2511.15830].

This framing distinguishes MAPs from benchmarks that isolate only one capability. In MAPs, park operation is not reduced to static layout design: the formal task includes pricing, inventory, staffing, research, and optional experimentation in a sandbox before evaluation. A plausible implication is that MAPs is meant to test whether an agent can integrate operational control, environment modelling, and strategic adaptation within one sequential decision process rather than solve these components independently.

Micro RollerCoaster Tycoon provides a complementary reference point. Its goal is to place rides and shops in an amusement park to maximize profit earned from park guests, and it was used to study whether starting from a minimal starting point for evolution and complexifying incrementally is beneficial, and what effects resource limitations have on creativity and optimization. This suggests a lineage in which amusement-park simulators moved from quality-diversity exploration of layouts toward broader tests of business-style decision making [2105.04342].

## 2. Formal environment model

In MAPs, the park state at day \(t\) is formalized as
\[
s_t=\bigl(O_t,R_t,S_t,T_t,\rho_t,m_t,\,\text{guest\_stats}_t\bigr),
\]
where \(O_t\in\{0,1\}^{20\times 20}\) is the occupancy matrix, \(R_t\) and \(S_t\) record each ride and shop’s subtype, subclass, location, price \(p_i\), stock \(q_i\), and dynamic stats, \(T_t\) is staff allocation and locations, \(\rho_t\in\{\text{“none”,“slow”,“medium”,“fast”}\}\) and topic queue encode research progress, \(m_t\in\mathbb{R}_+\) is money on hand, and \(\text{guest\_stats}_t=(\lambda_t,n_t,D_t)\) summarizes estimated arrival rate, current guest count, and aggregate satisfaction features [2511.15830].

Each morning the agent chooses exactly one action \(a_t\in A\). The action set includes:
- `place(x,y,type,subtype,subclass,price,order_quantity)`
- `move(id,x′,y′)`, `remove(id)`
- `modify(id,price)` or `modify(id,order_quantity)`
- `hire(subtype,subclass)`
- `set_research(topic,speed)`
- `survey_guests()`, `wait()`

The per-period reward is defined as operating revenue minus operating cost minus overhead:
\[
r_t = \sum_{i\in N_t} \bigl(p_i\cdot D_i(p_i)\bigr)\;-\;\sum_{i\in N_t} c_i(x_i)\;-\;C_{\mathrm{overhead}},
\]
and the finite-horizon objective is
\[
V = \sum_{t=1}^H r_t
\quad(\text{no discounting, }\gamma=1).
\]

Transition dynamics are stochastic. Guest arrivals follow a Poisson whose rate depends on park rating and capacity; breakdowns occur with per-ride Bernoulli probability \(\beta_i\); dirt accumulates proportionally to guest footfall; and staff cleaning and mechanics repair reduce dirt and downtime according to deterministic rates plus small noise. Spatially, the park is a \(20\times 20\) grid with a path matrix \(P\), water matrix \(W\), and occupancy matrix \(O\). Guests follow shortest-path distances on \(P\), and a ride’s appeal is increased by the number of adjacent water tiles. This combination makes the state partially operational, partially geometric, and explicitly dynamic [2511.15830].

## 3. Decision variables, difficulty modes, and sandboxing

MAPs exposes a daily business-control interface. The agent may choose locations \((x,y)\) for attractions or shops, set ticket prices \(p_i\in[\min_i,\max_i]\) for rides and item prices for shops, choose order quantity \(q_j\) for each shop to balance stockouts versus waste, allocate staff by hiring or firing mechanics, janitors, and specialists and placing them on the grid, and sequence research topics to unlock higher-tier entities [2511.15830].

The benchmark defines three official difficulty modes.

| Mode | Horizon and unlock state | Emphasis |
|---|---|---|
| Easy | \(H=50\) days; all ride/shop/staff subclasses unlocked; no research | Basic capacity-pricing trade-offs |
| Medium | \(H=100\); only yellow tier unlocked at start; research depth requires one or more research steps at monetary cost | Extended planning and research sequencing |
| Hard (in development) | \(H\approx 250\); additional mechanics include terraforming, guest-preference heterogeneity, and debt | Deeper long-horizon strategy |

In addition, every agent can optionally use a sandbox mode for up to 100 in-game days before evaluation. Sandbox actions are `max_money`, `unlock_all`, `reset`, and `switch_layout`. These actions allow curated experiments to learn transition dynamics under a limited sample budget.

A common simplification is to view amusement-park environments as only placement problems. MAPs is broader: placement is only one action family within a control regime that also includes pricing, inventory, staffing, research, surveying, and deliberate waiting. This is central to its use as a testbed for modelling business decisions [2511.15830].

## 4. Evaluation protocol and empirical baselines

The primary evaluation metric is final park value,
\[
V_{\mathrm{final}} = m_H + \sum_{i\in \mathrm{assets}} \mathrm{sell\_price}_i + \alpha\,(\text{IP from research}),
\]
where \(\alpha\) is a small constant for intellectual-property credit. Additional metrics are total profit \(\sum_t r_t\), customer satisfaction via average rides and shops visited, and sample efficiency in sandbox, measured as increase in \(V_{\mathrm{final}}\) per sandbox day [2511.15830].

Human players collectively achieved mean \(V_{\text{human}}=835\,318\) on easy and \(2\,062\,681\) on medium. LLM-based agents use a ReAct loop with a 5-step context window, conditioning on the JSON observation and the game manual. Evaluation is reported over three held-out layouts \(\times\) 3 seeds.

| Difficulty | GPT-5 Nano | GPT-5 | Grok 4 | Sonnet 4.5 | Gemini 2.5 |
|---|---:|---:|---:|---:|---:|
| Easy | \(0.36\%\pm0.28\) | \(13.89\%\pm11.16\) | \(6.62\%\pm14.50\) | \(4.53\%\pm5.75\) | \(3.96\%\pm8.23\) |
| Medium | \(0.64\%\pm0.78\) | \(7.16\%\pm8.36\) | \(0.62\%\pm0.39\) | \(1.43\%\pm1.40\) | \(0.64\%\pm0.29\) |

GPT-5’s easy-mode lead over the other LLMs is reported as \(p<0.05\) under a paired \(t\)-test across layouts. The aggregate comparison nevertheless remains strongly unfavorable to current systems: even the strongest model reaches only \(7.16\%\) of human performance on medium. This supports the benchmark’s claim that current agents struggle when open-ended objectives, temporal credit assignment, stochasticity, and spatial reasoning are coupled in one environment [2511.15830].

## 5. Diagnosed capability gaps

The MAPs evaluation isolates five recurring failure modes. First, under **open-ended objectives**, models exhibit myopic, greedily optimized strategies. Second, under **long-horizon planning**, doubling the horizon from \(50\rightarrow 100\) and adding research cuts GPT-5’s relative score from \(13.9\%\rightarrow 7.2\%\). Third, under **active world-model learning**, allowing 100 sandbox days with “undo” and “max_money” produces no consistent improvement: GPT-5 moves from \(13.9\%\rightarrow 15.5\%\) on easy but often degrades on medium, and qualitative analysis reports overfitting to day-specific observations, regurgitation of manual text, and failure to extract actionable, generalizable hypotheses [2511.15830].

Fourth, under **spatial reasoning**, a simple heuristic that places rides near water, shops at path intersections, and clusters attractions raises GPT-5 from \(13.9\%\rightarrow 21.6\%\) on easy, while LLM-only policies often produce parks that lack density or misinterpret winding path distances. Fifth, under **stochastic transitions**, per-day coefficient of variation for revenue can exceed \(0.15\) in early stages and drops to approximately \(0.05\) later. In the reported planning experiment, MPC with a learned WALL-E world model degrades performance because of prediction errors, whereas an oracle model yields \(>4\times\) improvement over the policy alone. The random-shooting MPC uses \(K=5\) rollouts and \(M=4\) world-model steps.

These findings also correct a possible misconception about sandbox access: optional experimentation does not by itself solve model-learning deficits. The published results indicate that better experimentation policies and better summarization of experimental outcomes are both necessary before sandboxing becomes reliably beneficial [2511.15830].

## 6. Relation to MicroRCT and implications for park-design research

Micro RollerCoaster Tycoon defines a park on a \(30\times 30\) tile grid with a single, “donut-shaped” main path, with the entrance in the upper-left, that is pre-built and immutable. Attractions may be added, removed, or replaced on empty non-path tiles, and when a new attraction is added, depth-first-search on the free-space graph builds the shortest connecting spur from the attraction’s door to the main path. The environment includes 24 attraction types, including 5 rollercoasters, thrill rides, transportation rides, cinemas, Circus, Crooked House, shops, restrooms, and first-aid; each has a cost per tile and metadata \(\{\text{excitement},\text{intensity},\text{nausea}\}\) matching original RCT values [2105.04342].

MicroRCT defines total profit as
\[
P = \sum_t \mathrm{Revenue}_t - \sum_a \mathrm{Cost}_a,
\]
and uses static and runtime descriptors including average excitement, intensity, nausea, Shannon-entropy ride-type diversity, happiness, vomit, and revenue. Its search procedure is MAP-Elites with a single objective, maximize total profit \(P\), and three 2-D behavior-characterization pairings: Excitement vs Intensity, Happiness vs Ride Diversity, and Happiness vs Vomit. Initialization is either small-park, with 0–4 attractions and mutation bias \(1/27\) remove versus \(26/27\) add/replace, or medium-park, with 8–12 attractions and 50/50 add versus remove. There is no crossover operator; only mutation is used [2105.04342].

The experimental design comprised \(2\) initial-size \(\times 2\) cost \(\times 3\) descriptor pairings \(=12\) experiment types, with 20 independent runs per type, 240 runs total, and 10,000 generations per run. Cost-disabled parks routinely reached \(\$50\text{k}–\$55\text{k}\) peaks, while cost-enabled parks were capped lower at approximately \(\$30\text{k}–\$40\text{k}\). For the Excitement-Intensity pairing, reported QD scores were \(\$27{,}801.56\) for small, cost enabled; \(\$37{,}198.09\) for small, no cost; \(\$26{,}776.50\) for medium, cost enabled; and \(\$35{,}975.36\) for medium, no cost. In every pairing, “small + no cost” yielded the highest QD score, and cost-disabled runs filled more cells in the MAP [2105.04342].

The reported elite maps further show that, in the Excitement-Intensity space, the highest-profit cells correspond to moderate excitement and intensity around 30–40 and use a mix of coasters and freefall rides; in the Happiness-Vomit space, the most profitable parks occupy the low-happiness, high-vomit region. Two distinct high-profit strategies are identified: “Shock & awe,” with many nauseating rides and minimal shops, and “Comfort funnel,” with fewer coasters and many concessions. A plausible implication for MAPs is that the amusement-park domain naturally supports multiple behaviorally distinct profit strategies, which makes it suitable both for quality-diversity analysis and for diagnosing failures in long-horizon business decision making.

## 7. Open directions

The MAPs roadmap includes a hard difficulty with \(H\approx 250\), terraforming, guest-preference heterogeneity, and debt. Additional proposed directions are richer observation modalities, including grid-vector inputs for traditional RL and image plus text for multimodal LLMs; improved active world-model learning through concise, general causal hypotheses; scalable stochastic world models through hybrid neuro-symbolic or programmatic inference beyond WALL-E; and spatial or graph-based planners that explicitly optimize layout density and flow by embedding \(P\) and \(W\) into a learnable graph representation [2511.15830].

Taken together, the published results position MAPs as a benchmark for holistic decision-making competence in a controlled but open-ended business environment, while MicroRCT demonstrates that even a simpler amusement-park simulator already exhibits wide design-space diversity and strong interactions between budget, layout, guest experience, and revenue. The combination of these two lines of work suggests that mini amusement parks are not merely game environments; they are compact laboratories for studying optimization under uncertainty, world modelling, and strategic adaptation [2511.15830]

Source: https://www.emergentmind.com/topics/mini-amusement-parks-maps