Titulo Estágio
Predicting Performance for Efficient Robotic Evolution
Local do Estágio
DEI-FCTUC
Enquadramento
Designing robotic controllers through evolutionary algorithms is a promising approach to develop robust and adaptable behaviours, particularly in tasks where hand-designed solutions are difficult to engineer. However, a major limitation is the high computational cost of evaluating candidate controllers through physics-based simulations or real-world experiments, especially in complex tasks such as chemical source localisation.
Typically, evolutionary processes rely on manually designed fitness functions to guide the search. These functions often introduce human bias and may not correlate well with real-world success criteria. Furthermore, evaluating each individual controller through multiple simulations to ensure performance reliability can be prohibitively expensive.
A promising alternative is to replace or augment the hand-crafted fitness function with a learned model (such as a neural network) that can predict a controller’s success and efficiency based on features extracted from a single run. This data-driven model would serve as a surrogate evaluator, reducing the number of simulations required and minimising design bias.
Objetivo
This dissertation aims to design, implement, and evaluate a machine learning-based system that can predict the performance of robotic controllers from a single evaluation, thereby significantly reducing the computational cost of the evolutionary process. The proposed approach focuses on tasks such as chemical source localisation, where repeated evaluations are expensive and time-consuming. The goal is to train a Machine Learning model (such as a Convolutional Neural Network) to estimate both the success rate and time efficiency of a controller by analysing a single trajectory of sensorimotor data. Its predictions will then be used to guide the evolutionary process, replacing or augmenting hand-designed fitness functions, in order to make it less biased and more sample-efficient. The system will be assessed by comparing the controllers evolved using this surrogate model against those obtained using traditional evaluation methods, with a focus on solution quality as well as convergence speed.
Plano de Trabalhos - Semestre 1
1 - Literature review
2 - Implementation of simulation environment and controller framework
3 - Data collection from simulated evaluations
4 - Design and training of the predictive model
5 - Writing of the intermediate report
Plano de Trabalhos - Semestre 2
6 - Integration of the learned model into the evolutionary loop
7 - Validation
8 - Writing of a scientific paper
9 - Dissertation writing and submission
Condições
The work will be conducted within the bio-inspired Artificial Intelligence (bAI) group from CISUC. There is a possibility of the student being awarded a scholarship (Bolsa de Investigação para Licenciado) for at least 6 months, renewable for an equal period by agreement between the advisor and the intern. The scholarship will follow the Fundação para a Ciência e Tecnologia (FCT) monthly stipend guidelines.
Orientador
João Macedo
jmacedo@dei.uc.pt 📩