Simulating Public Space Use using Agent-Based Models
This article presents ongoing research on spatial models that capture human activity in open public spaces. It contains work undertaken at the Centre for Advanced Spatial Analysis during my doctoral research. This article focuses on one aspect of the research, specifically the simulation of park activity using a particular set of modelling techniques, called Agent-Based Models (ABM). For a more detailed discussion on this work, readers are referred to the full text of the doctoral thesis, titled 'Agent-Based Modelling of Public Space Activity in Real-Time'.
Agent-Based Modelling (ABM) is a computational approach that represents systems through the interactions of their individual parts, called agents. Bonabeau captures the characteristics of ABM quite succinctly:
The benefits of ABM over other modeling techniques can be captured in three statements: (i) ABM captures emergent phenomena; (ii) ABM provides a natural description of a system; and (iii) ABM is flexible. It is clear, however, that the ability of ABM to deal with emergent phenomena is what drives the other benefits. (Bonabeau, 2002)
All three aspects mentioned above are highly relevant in the study of public space activity. Activity in open public spaces is mostly (if not exclusively) driven by the desires and interactions of individual people in the space, as there is no central regulator dictating how a space should work, in other words it is an emergent phenomenon. Therefore, when attempting to capture and replicate this activity, a framework that can support this natural representation is needed. Finally, as activity in public spaces is driven by individuals' needs and wants, it inherently accommodates a wide scope of behaviours, and therefore a framework would need to be sufficiently flexible to support this range of activities.
This work focusses on the scale of the individual person. At this scale, the physical environment around us is perceived through its full 3 dimensions, identified through the various facades, walls, windows, ledges, stairs, trees, urban equipment, and all other elements that make up public space. For this reason, the model presented here operates in a fully 3D environment. To accommodate for the third dimension, the model was developed in Unity, a 3D game development software that lends itself to dynamic simulations (Juliani et al. 2018). The rest of this article will discuss the workings of the agent-based model developed here.
This work focusses on the scale of the individual person. At this scale, the physical environment around us is perceived through its full 3 dimensions, identified through the various facades, walls, windows, ledges, stairs, trees, urban equipment, and all other elements that make up public space. For this reason, the model presented here operates in a fully 3D environment. To accommodate for the third dimension, the model was developed in Unity, a 3D game development software that lends itself to dynamic simulations (Juliani et al. 2018). The rest of this article will discuss the workings of the agent-based model developed here.
|
Each agent in the model represents one park visitor. An agent has a physical presence in the model, with its avatar modelled as a simple volume with dimensions 1 x 0.5 x 2 (width x depth x height, in metres). Furthermore, agents have a basic set of cognitive skills that simulate vision which allow them to detect the physical environment around them, and enable them to navigate within the area.
In addition to their physical characteristics, agents have a set of additional behaviours, called activities, which they can engage in within their environment. The four activities are Wander, Sit, Feature Visit, and Sports, and are presented in more detail later. Each agent is introduced in the simulation with a predetermined visit duration, during which it wanders the area and engages in activities at random. At the end of each agent's visit duration, it moves to exit the area and is removed from the simulation. |
An agent avatar.
|
|
Agents are able to move within the environment using a Wander behaviour. The wandering behaviour is implemented using an angular constrained random walk, where an agent selects a location at random within their vision range and angle of view, and starts moving towards that location. Once the destination is reached, they may choose a new location to move. Walking is the default agent activity available to all agents, as it allows them to navigate the environment and enables the execution of all other activities.
|
Animation of agent random walk
(Click for larger version)
|
|
The first stationary activity is a Feature Visitor activity. When an agent chooses to engage in this activity, they select a location from a predetermined list that represent fixed-point attractions in the area (i.e. playgrounds, restaurants, etc). If there are no attractions in the area of interest, this activity is not available. Once an agent has selected a valid target attraction, they calculate the shortest path to it, and start moving towards it. Once the destination is reached, they situate themselves near it and stay stationary for a fixed period. At the end of the period, the agent chooses a new activity at random.
|
Animation of agent feature visitor activity (Click for larger version)
|
|
The second stationary activity is a Sports activity. When an agent chooses to engage in a Sports activity, they immediately enter a loop during which they start searching for a valid location within their local environment at which they can engage in a Sports activity.
The conditions for a valid location are set so that the area is clear of obstacles (trees, buildings, other stationary agents not in Sports state) within a certain radius (representing the play area), and it is on flat terrain. While searching for a valid location, the agent samples locations around them at random, until a location satisfying all the above conditions is found. While sampling, they implement a continuous wandering behaviour, ensuring that they cover a large area in their search. When a valid location is found, the agent moves to that location, and remains stationary for a fixed period. At the end of the period, they choose a new activity at random. |
Animation of agent sports activity
(Click for larger version)
|
|
The third and final activity is a Sit activity. It is a generic activity, and represents all social activities that might take place in a park that are not tied to a specific location. When an agent chooses to engage in a sit activity, they immediately start a searching behaviour, similar to the Sports activity, during which they look for the optimal location to engage in a Sit activity. In contrast to Sports, for a Sit activity the agent samples nearby locations for a predetermined duration, during which it implements a wandering behaviour, allowing it to cover a larger area. During sampling, the agent scans random locations within its vision that are on valid terrain (not in water or paths), scoring each location by the number of other agents visible from there (more agents visible = higher score). Furthermore a restriction is imposed, by which a potential location is discarded if another agent is occupying a spot within a small distance (less than 5m) of it, ensuring that agents do not pile on top of each other. At the end of the sampling period, the agent chooses the location with the highest score, moves there, and remains stationary for a predetermined amount of time. At the end of the Sit period, they choose a new activity at random.
|
Animation of agent sit activity
(Click for larger version)
|
To test the model, a sample park environment was developed, containing open green areas, paths, water features, feature attractions, and trees. Additionally, specific points in the park perimeter were annotated as park gates, signifying the locations where agents may enter and exit the park.
 Environment layers
|
Final 3D model for the environment.
|
A agent behaviour framework was tested in the sample park environment by executing multiple simulation runs. A simulation run is presented here, with agents exhibiting all four behaviours. Behaviour probabilities were set equally for all agents, so that each time an agent picked a new activity to engage in, each activity had a specific probability to be picked and was furthermore colour-coded, as given in the following table:
| Activity | Probability | Colour |
|---|---|---|
| Sit | 0.2 | Blue |
| Feature Visit | 0.1 | Yellow |
| Sports | 0.1 | Red |
| Walk | 0.6 | White |
The simulation run can be seen in the video below, presented in two different views. The same simulation run is depicted in both views. The left screen follows a single agent throughout its visit, illustrating internal model mechanics. The right screen provides a top down view of the simulation, with real-time statistics and 3D bar charts of the spatial distribution of activities, colour coded by activity.
The simulation was run for 5000 steps, allowing for a stable state to be reached in the model. Spatial distribution of activity was measured using a regular grid of 50m cell size, which visualises cumulative relative activity in each cell. Bar heights are accurate for relative measures, so that different locations and activities in the same run can be compared.
The picture of overall activity shown in the simulation above shows some clear trends: Feature Visit activities (in yellow) are exclusively concentrated around the two attractors (bottom left, bottom right). Hotspots of sports activities (in red) are seen in multiple locations, with some clusters presenting a denser situation than others (bottom middle, left). Walk activity (in white) appears to be higher around paths and bottlenecks (middle). Sit activity (in blue) does not present any significant hotspots, but is scattered throughout the area.
Given the agent behaviour parameters discussed earlier, this activity is expected. By running additional simulations with limited available activities, we can investigate the effect each activity has on its own to overall use, as well as the effect one activity can have on another. In the following section, simulations were run with limited agent activities; each simulation was allowed to run for 5000 steps.
The picture of overall activity shown in the simulation above shows some clear trends: Feature Visit activities (in yellow) are exclusively concentrated around the two attractors (bottom left, bottom right). Hotspots of sports activities (in red) are seen in multiple locations, with some clusters presenting a denser situation than others (bottom middle, left). Walk activity (in white) appears to be higher around paths and bottlenecks (middle). Sit activity (in blue) does not present any significant hotspots, but is scattered throughout the area.
Given the agent behaviour parameters discussed earlier, this activity is expected. By running additional simulations with limited available activities, we can investigate the effect each activity has on its own to overall use, as well as the effect one activity can have on another. In the following section, simulations were run with limited agent activities; each simulation was allowed to run for 5000 steps.
A Walk only simulation run helps establish a baseline of activity in an area. With agents only allowed to wander, busy pathways may be identified as dictated by morphology. These are identified here as the bottlenecks imposed by the water features (center, bottom middle), as well as slightly increased activity on paths, as paths are the preferred terrain for moving.
|
|
As Feature Visit activities can only take place at specific locations, the only hotspots of activity are naturally centered around these predetermined locations. Increasing the Feature Visit probability parameter only affects the intensity of these hotspots.
|
|
Sports activity is driven primarily by the environment layout, as it is an activity highly dependent on morphology. Therefore agents quickly identify all the appropriate locations, and any increase in Sports probability only increases the hotspot intensity. It is interesting to note however that the denser Sports locations are visible from busy pathways, suggesting that visibility plays a significant role for increased activity.
|
|
The Sit activity is the most complex activity, as it essentially relies on the locations of other agents at all times. In general, some hotspots can be identified, close to busy intersections (center, center right), as well as locations that have a view to the busiest paths (around the river). Increasing the Sit probability increases intensity of hotspots at first, up to a saturation point, at which point other hotspots start to appear, and a spill-over effect can be observed.
|
|
By investigating activities individually, it becomes clearer how the overall activity emerges when all activities are available. Looking back the the simulation run of all activities: The fixed locations of Feature Visits attract Feature Visitors which generate hotspots at those locations. Locations suitable for Sports activities are quickly taken up by Sports agents, discouraging other activities within them (such as Sit activities). Sit activities are found around busy pathways similar to a Sit-only run, but in this case some clusters of Sit activity are observed close to hotspots of other activities. This is expected: Although the open areas are taken over by Sports, the increased agent presence there encourages minor Sit activity in its peripheries; a similar effect is observed with the Feature Visit locations, which drive up neighbouring Sit activities.
The model was evaluated through two case studies, where visitor behaviour for two real-world parks was simulated. The simulation results were then compared to observed visitor activity, captured through on-site surveys. Simulation results in both case studies are shown in two ways: In real-time during the simulation, a regular grid is overlaid over the whole area, and both simulated and observed visitors in each cell are counted as a proportion of the total respective population. 3D bars are then used to plot the error between the two relative values, colour-coded red for model overestimation (the model predicted more visitors in that location), and blue for model underestimation (the model predicted less visitors in that location). For subsequent analysis, at a specific time in the simulation a snapshot is taken of all the agent locations at that point in time. This spatial distribution is then plotted as a heatmap which can be compared against a heatmap of observed visitor locations, as shown in the maps. The 'Layers' button in the top-right corner of the maps allows toggling between the two datasets.
|
The first case study focussed on Hyde Park in London, UK. Hyde Park presents a fairly flat terrain, with a single major feature being a river that runs through it almost completely. Activity in Hyde Park appears to be centered around three locations, as seen in the map below. Each of these locations contains an attractor (a cafe and fountain, a restaurant, and a designated public speaking spot at bottom left, center, and top right, respectively), which seem to be the driving factors for the observed activity. This is implemented in the model using Feature Visit targets at these locations which capture the general position of hotspots, but fail to replicate the spread of activity found around the hotspots. Otherwise, the model performs well in replicating activity and lack of it at other locations in the area.
|
|
|
The second case study focussed on Queen Elizabeth Olympic Park (QEOP) in London, UK. QEOP is a newly developed park, designed for the London Olympic Games in 2012, and opened to the public soon afterwards. It has a more varied terrain compared to Hyde Park with noticeable height differences, and the river Lea and canals running through it, splitting it to multiple land areas connected via bridges. Similar to Hyde Park, most of the activity is located around attractors of activity, namely 2 restaurants and a playground.The model performs well at capturing this activity using Feature Visit targets, but overestimates sports activities, mainly towards the South end (bottom right in the model). Furthermore, a wide pathway connecting the park to a major retail park to the East attracts significant footfall and stationary activities, which is underestimated in the model. The reason for this underestimation is the proximity of the pathway to the border of the model environment, which discourages activities taking place near the edges, as it has been developed for continuous spaces.
|
|
|
Share:
|
|