Prometheus started off built around training neural networks in a competitive environment to emulate behaviors of organic beings. We wanted to explore the role of algorithms in society, the transactional relationship between the cityscape and the natural world.
Work completed in semester one consisted of training machine learning agents and developing environment assets such as buildings and terrain. A few prototypes were made of agents trained to seek out a resource and then push the resource into a goal point.
However, we found training through the Unity ML Agents package was rather difficult to get the agents to do what we wanted, most of the time they would not learn or become too accustomed to the training environment (overfitting). After a semester of work we could see it was going to be hard to achieve our goals of emergent ML agent behavior in Unity with the current tools. Perhaps in the future we could return to exploring this avenue through using the public tools provided by OpenAI to create their hide and seek demo. We found that the conceptual themes we wanted to explore could be addressed without the use of machine learning agents.
As the project stands, we shifted our attention from machine learning to simpler scripted agents to produce the same concept. The time sensitive nature of this project made us question the things that are really important in terms of our final depiction.
Building on top of this we moved onto an agent vs agent environment where two agents on opposite teams competed for the same resource, similar to soccer.
A test of the rotating camera angle and larger environment.
This milestone saw a significant jump in the visual quality of the environment, which began to take shape and become a relatively polished scene.
Significant terrain forming and other aesthetic changes.
This week saw the introduction of multi-faction competition and randomization of building placement.