Research blog posts from the Foerster Lab for AI Research (FLAIR) at the University of Oxford
Latest publications and findings from our lab
AH2AC2 is a benchmark for evaluating how well AI agents can coordinate with humans in the complex, partially observable, cooperative game Hanabi, especially when access to human data is limited.
Continue Reading →In Part I of this blog, we characterised the stability of TD through the TD Jacobian. In this part, we now build on this analysis to better understand the reasons for instability before proposing a surprisingly simple architectural solution that can stabilise TD.
Continue Reading →In this first blog of a two-part series, we take a deep dive into understanding the challenges faced when developing stable TD algorithms and introduce a powerful tool to formally characterise TD's instability mathematically.
Continue Reading →We prompt the LLM to repeat its own output, which protects the model from adversarial attacks by avoiding the auto-regressive trap. We name this method PARDEN. PARDEN is particularly effective in the relevant regime of high True Positive Rate and low False Positive Rate.
Continue Reading →In summary, we show that minimax regret has a notable failure case when there are environments with high irreducible regret. Our solution concept can address this problem, and ReMiDi results in higher empirical performance in cases like these.
Continue Reading →We present JaxMARL, a library of multi-agent reinforcement learning (MARL) environments and algorithms based on end-to-end GPU acceleration that achieves up to 12500x speedups.
Continue Reading →