Abstract:This research work aims to develop an analytical approach for optimizing team formation and predicting team performance in a competitive environment based on data on the competitors' skills prior to the team formation. There are several approaches in scientific literature to optimize and predict a team's performance. However, most studies employ fine-grained skill statistics of the individual members or constraints such as teams with a set group of members. Currently, no research tackles the highly constrained domain of the FIRST Robotics Competition. This research effort aims to fill this gap by providing an analytical method for optimizing and predicting team performance in a competitive environment while allowing these constraints and only using metrics on previous team performance, not on each individual member's performance. We apply our method to the drafting process of the FIRST Robotics competition, a domain in which the skills change year-over-year, team members change throughout the season, each match only has a superficial set of statistics, and alliance formation is key to competitive success. First, we develop a method that could extrapolate individual members' performance based on overall team performance. An alliance optimization algorithm is developed to optimize team formation and a deep neural network model is trained to predict the winning team, both using highly post-processed real-world data. Our method is able to successfully extract individual members' metrics from overall team statistics, form competitive teams, and predict the winning team with 84.08% accuracy.

AutoCompete: A Framework for Machine Learning Competition

Benchmarking Automatic Machine Learning Frameworks

Multi-Objective Automatic Machine Learning with AutoxgboostMC

Techniques for Automated Machine Learning

AMLB: an AutoML Benchmark

A Neophyte With AutoML: Evaluating the Promises of Automatic Machine Learning Tools

Online AutoML: an adaptive AutoML framework for online learning

AutoKaggle: A Multi-Agent Framework for Autonomous Data Science Competitions

Winning Solutions and Post-Challenge Analyses of the ChaLearn AutoDL Challenge 2019

Automated Machine Learning: From Principles to Practices

A General Recipe for Automated Machine Learning in Practice

Benchmark and Survey of Automated Machine Learning Frameworks

Automated Imbalanced Learning

Ensemble Squared: A Meta AutoML System

AutoDES: AutoML Pipeline Generation of Classification with Dynamic Ensemble Strategy Selection

Automated machine learning with dynamic ensemble selection

Towards robust and domain agnostic reinforcement learning competitions

AutoML @ NeurIPS 2018 challenge: Design and Results

An Integrated Framework for Team Formation and Winner Prediction in the FIRST Robotics Competition: Model, Algorithm, and Analysis

Analysis of systems' performance in natural language processing competitions

Taking Human out of Learning Applications: A Survey on Automated Machine Learning.