An agent-based model (Mesa) exploring how forestry intensity shapes herbivore access to protected reserve land — built around the research context of SLU's Grimsö Wildlife Research Station.
Grimsö's research explicitly studies how land-use practices — forestry, agriculture, infrastructure — affect the space use and population dynamics of wildlife such as moose and roe deer. This project models a simplified version of that relationship: how the proportion of forest under active logging affects where a herbivore population spends its time, and how easily it can reach protected reserve patches.
This is the third of three linked portfolio projects submitted for the SLU Research Engineer (AI-driven Social Simulations) application: a land-use negotiation simulator, statistical analysis of negotiation outcomes, and — here — what happens to wildlife on the ground once a land-use decision is made.
- Grid: 20×20 Mesa
MultiGrid. Cells are typed asforest,logged, orreserve. - Reserve: a single contiguous block (~15% of the grid) in one corner, mirroring a real bounded protected area.
- Forestry intensity: the proportion of the remaining (non-reserve) land actively logged, tested at 20%, 50%, and 80%.
- Herbivore agents: each step, an agent looks at its 8 neighbouring cells and moves to whichever scores best — reserve cells are favoured, logged cells are penalised, forage level also counts. It then eats forage from its current cell.
- Forage: every cell has a forage value that regenerates a little each step, more slowly on logged cells.
Reserve-dependency falls as forestry intensity rises — the opposite of the project's original hypothesis, and a more interesting finding.
| Forestry intensity | Mean reserve-dependency (10 replicates) |
|---|---|
| 20% | 0.92 |
| 50% | 0.65 |
| 80% | 0.51 |
This is driven by habitat fragmentation, not resource avoidance. Herbivores only ever assess their immediate neighbourhood, so they navigate the landscape locally, one step at a time. At low forestry intensity, forest cover stays largely connected, so herbivores can find a route to the reserve fairly easily. At high intensity, remaining forest becomes fragmented into small, disconnected pockets — herbivores can get functionally stranded, surrounded by logged land in every direction, unable to find a path to the reserve even though they are actively trying to avoid logged cells at every step.
This mirrors a well-documented real-world ecological effect: fragmentation can restrict access to refugia independently of overall habitat quality.
Final reserve-dependency vs. forestry intensity, all replicates shown, with the mean trend line.
Reserve-dependency over the full 150-step run for one representative replicate per intensity level, showing that higher intensity slows and dampens the settling process, not just its endpoint.
Spatial snapshot of the landscape and herbivore positions after 150 steps, for all three intensity levels side by side. Makes the fragmentation mechanism directly visible.
python3 -m venv venv
source venv/bin/activate
pip install mesa pandas matplotlib networkx
python sweep.py # runs the full parameter sweep, saves CSVs
python plot_results.py # produces figure1 and figure2
python landscape_snapshot.py # produces figure3- Applied agent-based modelling with Mesa: environment design, agent behaviour rules, and emergent population-level patterns.
- Translating a real research question — the effect of forestry intensity on wildlife space use — into a testable computational model.
- Parameter-sweep experimental design with replicates, and honest, iterative interpretation of simulation output (the original hypothesis was revised after the data showed a clearer, opposite pattern).
- Direct familiarity with the host department's actual research context and priorities.
Predator–prey dynamics (wolf/lynx interaction with herbivores) are a natural next step and are also core to Grimsö's research, but were intentionally left out of this build to keep scope realistic within a short build window. A natural extension would add a predator agent type and examine how predation risk interacts with the fragmentation effect already observed here — for example, whether herbivores in fragmented, high-logging landscapes become more vulnerable to predation due to reduced access to safe reserve areas.