EcoMeet

Inspiration

• We used multi agent reinforcement learning as we did not have test data, we wanted to optimise using a number of parameters for multiple agents dependent on each other.
• Dynamic visualisation to show the user the specific new predicted

What it does

• The environment learned a policy using PPO, to maximise the reward which included factors such as CO2 emissions, travel time, availability
• The action space are the airports (including the offices)
• Model compiled and outputs the optimal destination dependent on the factors at current state.
• Front end shows flight paths, 3D globe visualisation with markers for QRT offices as the initial state and output is model response in output JSON format.
• Backend handles the post request of input JSON, preprocesses the JSON and then passes into the model, then constructs the output JSON including the exact coordinate of the airport.,

How we built it

• Model: PettingZoo multi-agent reinforcement learning environment, stablebaselines3,
• Backend: Complied model with FastAPI
• Frontend: React, Tailwind, Cesium (3D Globe), Blender (QRT Plane Model)

Challenges we ran into

• Challenge 1: Used wrong approach for RL, single agent instead of multi-agent initially
• Challenge 2: Segmenting training data into trainable chunks
• Challenge 3: RL Model training time, inefficiency in caching
• Challenge 4: Optimisation of model response
• Challenge 5: Making custom 3D model of plane

Accomplishments that we're proud of

• Stylistic and dynamic frontend
• Accurate model
• Efficient training and outputting

What we learned

• Handling multi-agent reinforcement learning
• Training with Big Data
• 3D modelling
• FastAPI
• AI Agentic tooling

What's next for EcoMeet

• With more time and resources we could have utilised the big data pool
• More factors E.g. Integrating with the API

Built With

• blender
• cesium
• fastapi
• gemini
• pettingzoo
• python
• react
• stablebaselines3
• tailwindcss