Ai-assistant for Architects at early design stage
Following work is envisioned as a reflection on the design principles and methodologies learned during my studies as an architect. From my early days in architecture, I was fascinated by the conceptual phase of project development. I found it truly amazing how diverse design inputs could merge into something completely new. My intention for this work was to experiment with the traditional architectural workflow by abstracting it into key ideas and present visions of how it could be transformed.
This thesis explores the integration of agent-based modeling into architectural design, aiming to enhance the early design stage through a flow-driven approach. It proposes a method where intelligent agents, mimicking processes of natural morphogenesis, aid in data analysis and decision-making. The aim of the resulting tool is to make work with data more interactive by enabling architects to communicate with data through agents.
Stage 1 Workflow Analysis
In traditional architectectural workflow architect’s vision was the only tool for synthesizing various design inputs, that in this work will be refered as flows. Flows are input parameters that form a concept of a space before its physical definition is determined.
![](https://blog.iaac.net/wp-content/uploads/2024/06/123.gif)
In different periods of time depending on how developed our instruments were, architects were giving preference to one of the flows. Until very now when we reached the point when we have enough computation power to process all the flows simultaneously, which means that today’s architecture can be flow-driven.
![](https://blog.iaac.net/wp-content/uploads/2024/06/3-12.jpg)
Past decades softwares were integrated into architect’s workflow. This caused a shift of the peak of decision-making process from detailed design stage to early design stage.
![](https://blog.iaac.net/wp-content/uploads/2024/06/5-6.jpg)
With a larger amount of data now being available to architects during the early design stage, the abundance of imputs has made the exploration process more challenging. It increased the importance of early design decisions by creating a point after which, the effectivness of decisions decrease, while the impact and possibility of change increaes. Architects can no longer depend solely on their intuition and expertise to make informed decisions. They need a methodology to manage and interpret this data in their designs.
![](https://blog.iaac.net/wp-content/uploads/2024/06/6-6.jpg)
The approach used in practice for making sense out of the project data implements optimization and simulation techniques. The gap that could be outlined in it is that the whole procedure of refining the input parameters requires primary conclusions to be made beforehead, thus making architect still the one responsible for analyzing the data and decision-making.
![](https://blog.iaac.net/wp-content/uploads/2024/06/2-10.jpg)
Nature has an ultimate ability to compute complex flows and their outcomes. If we compare architectural workflow with natural morhoghenesis, its components would appear quite similar, however, the computing process would be very different. In nature signals are not directly transfered into commands, there is an intermidiate step that allows to process them before coming up with response. This step has unique components that work together in cycles. What if we enhance architectural workflow with new nature-like behaving components?
![](https://blog.iaac.net/wp-content/uploads/2024/06/7-7.jpg)
Agent-based modeling was chosen because it excels at simulating complex systems, much like those found in nature. In nature, morphogenesis is driven by the interactions of numerous autonomous agents (cells) that follow simple rules yet collectively result in highly organized structures. Similarly, architecture involves a multitude of complex inputs—such as environmental data, material properties, and user requirements—that interact in intricate ways to shape the final design. By using agent-based modeling, we can effectively mimic the natural processes of morphogenesis in an architectural context. This methodology allows us to create a virtual environment where autonomous agents, each with specific behaviors and goals, dynamically interact with their surroundings and with each other.
![](https://blog.iaac.net/wp-content/uploads/2024/06/8-6.jpg)
The goal is to provide architects with an ai-assistant based on agents system that will allow interactive process of designing and support architect in desicion-making process at early design stage.
![](https://blog.iaac.net/wp-content/uploads/2024/06/9-5.jpg)
Stage 2 Algorithm Development
The integration of agents into the architectural workflow is designed to create an interactive and dynamic process where agents serve as intermediaries between architects and data. This integration occurs at the early stages of the design process, allowing architects to engage directly with design inputs without pre-formed conclusions. Unlike traditional tools that offer solutions based on defined objectives, this approach enables architects to explore data concurrently with the design process. The case study was conducted in an abstract manner: agents were associated with spaces and their definitions such as activities and their requirements. Activities were selected as defining factors because they contribute to all the flows within the system: space use as information flow, conditions as energy flow, spatial limits as matter flow.
![](https://blog.iaac.net/wp-content/uploads/2024/06/1-4.gif)
initializing – input data – setting up an environment
![](https://blog.iaac.net/wp-content/uploads/2024/06/hhhhhhhhhhhhh-1.gif)
project data – input parameters – setting up agents
Agents are autonomous entities of various types, following a set of rules, endowed with limited memory and cognitive ability, embedded in a network of other agents within the environment. Each agent in the model has its own set of characteristics and behaviors, which can change in response to interactions with other agents and the environment. In the developed system agents are categorized into two types in analogy with the discovered functions in morphogenesis: there are sensing agents and processing agents. Sensing agents are tasked with exploring the environment, making observations and creating a field of options. These agents are dependent, meaning their primary function is to study the environment, but they are not capable of decision making. Their role in the system is to guide processing agents. Processing
agents in their turn are responsible for distributing themselves within the environment based on the options field created by sensing agents. Each processing agent has a set of associated sensing agents. But unlike sensing agents that consider their individual requirements only, processing agents are linked to the project documents, from where they can seek additional information to consider in the decision-making process. Together, they dynamically interact with the environment. Architects can influence this process, by changing the input data, system variables or adjusting system rules, or they can send direct commands as a new input to the processing agents.
![](https://blog.iaac.net/wp-content/uploads/2024/06/jljlk.jpg)
types of agents
![](https://blog.iaac.net/wp-content/uploads/2024/06/223567.jpg)
relationships between agents and the environment
![](https://blog.iaac.net/wp-content/uploads/2024/06/Copy-of-ANIMATION.gif)
project data – input parameters – setting up agents
Agents were embedded with information about activity requirements. Sensing agents were programmed with the specific requirements of activities, representing various actions performed by individuals within residential spaces. Processing agents were linked to project data containing a list of spaces, where these activities occur. Activities were correlated with spaces for both individual and group activities, as well as private and shared spaces. Spaces were categorized into groups based on their use, while activities were scored based on their frequency, intensity, and duration, and then assigned a weight. Activities were further categorized as long-term or short-term and sorted based on their need for daylight.
![](https://blog.iaac.net/wp-content/uploads/2024/06/224.jpg)
project data – input parameters – setting up agents
_test 1. Input data corelation
Algorithm was tested in grasshopper enviroment. User interface was made with Human plugin to illustrate how input project data can be embeded in script as well as how the working parameters are related.
![](https://blog.iaac.net/wp-content/uploads/2024/06/fgnxxnnn.gif)
system set up
![](https://blog.iaac.net/wp-content/uploads/2024/06/Copy-of-Copy-of-Anastasija-Vidovic-copy-2.gif)
test with default settings
![](https://blog.iaac.net/wp-content/uploads/2024/06/dddddddddddddd.gif)
![](https://blog.iaac.net/wp-content/uploads/2024/06/21-2.jpg)
System rules in ABM are the set of principles, guidelines, or constraints that govern the behavior of agents and their interactions within the environment. These rules dictate how agents respond to each other and to changes in their surroundings. In the case of the developed algorithm, system rules are represented by decision-making criteria, which specify how agents make choices based on their requirements, relations and interactions with other agents and the environment. Key variables influencing agent behavior include the direction of movement of sensing agents, the radius of processing agents, and vision cone parameters.
![](https://blog.iaac.net/wp-content/uploads/2024/06/final-final-final.gif)
overview of availible tabs and control options
![](https://blog.iaac.net/wp-content/uploads/2024/06/23-1.jpg)
![](https://blog.iaac.net/wp-content/uploads/2024/06/мииииииииии.gif)
overview of availible working modes
A flow boundary is an emergent output of the system that is formed in collaboration of agents and architect. It visualizes the exploration process, focusing on spatial distribution and flow performance, thus helping architects assess and verify their design decisions. The tool’s primary aim is to assist in the exploration process, not to dictate solutions.
Stage 3 AI-ASSISTANT
The work is still in progress, more materials will be uploaded soon…
Currently I am working on integrating reinforcement learning (RL) into a designed ABM system to enables agents to learn through trial-and-error interactions with the environment, instead of only following system rules. Another way of enhancing the system that is under development is integrating an LLM to make this tool more user-friendly and allow more intuitive interaction. With LLM agents would be able to search for requested parameters and potentially perform tasks based on textual commands. Finally to further test the adaptability and robustness of the system another experiment is planned. It would involve selecting several inputs from various flows and running the model to observe its performance across a broader spectrum of data. This will include testing the system on real documents by taking on an actual design task and going through the entire design process.