· Supercell, Jätkäsaarenlaituri 1, 00180 Helsinki
An evening of talks at Supercell: reinforcement learning for game balancing, sleep and activity research with wearables, and lightning talks on networking apps, estimating car weight, and Marimo notebooks. And as always, food, drinks, and networking.
Sponsored by Supercell.
Heikki Arponen from Supercell presents how reinforcement learning is used to balance mo.co, Supercell's newest multiplayer PvE game. Learn how thousands of AI bots are trained to test new gear, content, and game features before release, helping designers iterate quickly on game balance.
Topics: building a headless game simulator that runs 1,000+ parallel environments per GPU; training recurrent neural networks with PPO (Proximal Policy Optimization); reward shaping challenges and the credit assignment problem; discovering game exploits through AI testing (like the 1-second dungeon completion bug); exploration techniques including k-nearest neighbor state entropy maximization; genetic algorithms for gear optimization; the "squid blades" problem — when RL agents can't learn stealth strategies.
Heikki shares practical insights from 20 months of RL development at Supercell, including what worked (position-based exploration), what didn't (latent space k-NN), and ongoing challenges in creating a generalizable RL solution for game testing.
Chandreyee Roy from Aalto University presents research on chronotypes (owls, larks, and neither types) using wearable devices to objectively measure sleep patterns and activity levels.
Topics covered: what are chronotypes and why they matter; limitations of traditional questionnaires (recall bias); mid-sleep time as chronotype indicator; sleep quality analysis using entropy and hypnograms; activity tracking differences between chronotypes; stress reporting patterns; wearable device comparison and user experience; future research on gender differences, body temperature, and objective chronotype detection.
Morningness-Eveningness Questionnaire: https://qxmd.com/calculate/calculator_829/morningness-eveningness-questionnaire-meq
Allan Nevala demonstrates a prototype app that helps you find the most compatible people to meet at networking events. Using AI embeddings and community detection algorithms, the system analyses attendee profiles to identify your best matches based on shared interests, location, and lifestyle compatibility.
In this 10-minute lightning talk, Allan walks through: profile matching using OpenAI embeddings; AI-generated conversation starters and meeting suggestions; community detection to identify interest clusters at events; an analytics dashboard showing compatibility algorithms; and privacy controls for safe networking.
Kevin demonstrates how to estimate car kerb weight using only dimensional measurements through supervised machine learning. He compares three approaches: scikit-learn's Random Forest, TensorFlow, and PyTorch neural networks.
Using data from Open Canada, Kevin builds models that predict car weight from features like overall length, width, height, and front-end length. He explains his neural network architecture (256-128-64 neurons), choice of ReLU activation, Adam optimizer, and why he uses Mean Absolute Percentage Error (MAPE) instead of MSE for this problem.
Code repository: https://github.com/axle217/MLexercise_CarDimensions
Teemu Säilynoja presents Marimo, a modern alternative to Jupyter Notebook that solves common pain points in interactive Python development.
Highlights: reactive execution, where cells automatically re-run when dependencies change; git-friendly notebooks saved as pure Python files, not JSON; no more out-of-sync cells, since the dependency graph ensures consistency; app mode to instantly turn notebooks into interactive web apps; LLM-friendly clean Python code that AI tools can easily read and fix; and native database integrations and testing support.
Perfect for data scientists and Python developers frustrated with traditional notebook limitations. Marimo brings modern development practices to exploratory data analysis while maintaining the interactive coding experience you love.
More about Marimo: https://marimo.io/
This event has taken place. The original event listing is on Meetup.