AI has recently shown impressive advances, from learning to play the Atari games to defeating expert human players in the game of Go. Beyond games, AI has also exploded in fields such as computer vision and natural language processing, where vast amounts of labeled data are available. More broadly speaking, technology as a whole has massively changed the landscape of most fields. However, current approaches can only help in tasks where we either can precisely specify the objective or already have plenty of observations of solutions to learn from.
Unfortunately, important real-world problems rarely have well-specified objectives or solutions to learn from. Instead, most problems depend on the goals and preferences of humans - the users - who are solving them. As a result, we need approaches that explicitly consider the user. We, the Artificial agents with Theory Of Mind team of FCAI, do exactly that by developing techniques and methods that assist users in their tasks.
For this we have developed a tutorial where you can see and play around with a concrete application of AI-assistance. Below you find a sample of our work.
This paper presents the general AI-assisted Design framework. Based on the recent rise in the use of AI in design practice, and on gaps that remain in current assistance systems for design, it argues for a new assistance paradigm: AI-assisted Design. We focus particularly on the experience of the designer, explaining why designer autonomy, as well as control over the AI assistant, is important for ownership and effectiveness. We explain the general outlines of the AI-assisted design framework and go in depth on some of the more novel aspect, including the need for user modelling and how one can ensure designer control.
De Peuter, S., Oulasvirta, A., & Kaski, S. (2023). Toward AI assistants that let designers design. AI Magazine, 44(1), 85-96.
Probabilistic user modeling is essential for building machine learning systems in the ubiquitous cases with humans in the loop. However, modern advanced user models, often designed as cognitive behavior simulators, are incompatible with modern machine learning pipelines and computationally prohibitive for most practical applications. In this work, we address this problem by introducing widely-applicable end-to-end differentiable surrogates for bypassing this computational bottleneck; the surrogates enable inference with modern cognitive models in modern machine learning with online computational cost.
Alex Hämäläinen, Mustafa Mert Çelikok, and Samuel Kaski. Differentiable user models. The 39th Conference on Uncertainty in Artificial Intelligence, 2023.
This paper presents a general-purpose instance of the AI-assisted design framework. We formulate an instance in which an assistant helps an agent (f.ex. a human) solve a decision problem through advice. We focus particularly on the negative effect of biases within the agent on the effectiveness of advice, and show that modelling biases can help mitigate these effects. Finally, we introduce an MCTS-based planning algorithm for finding the assistant’s optimal advice policy.
De Peuter, S., & Kaski, S. (2022). Zero-Shot Assistance in Sequential Decision Problems.
The paper describes a cooperative Bayesian Optimization problem where two agents work together to optimize black-box functions of two variables. This approach is motivated by collaboration between humans and AI, where an AI-assistant helps a human user solve a problem through collaborative optimization. The solution involves strategic planning of queries using Bayes Adaptive Monte Carlo planning and a user model that accounts for conservative belief updates and exploratory sampling of points to query. The paper presents a promising approach to cooperative optimization that has practical applications in human-AI teamwork.
Ali Khoshvishkaie, Petrus Mikkola, Pierre-Alexandre Murena, Samuel Kaski. Cooperative Bayesian optimization for imperfect agents. ECML-PKDD 2023