Making AI Better by Making it Irrational

The new technique predicts human error, by understanding irrationality, rationally. This technique could lead to better AI – human interactions.

Understanding and modelling human behaviour is a complex but crucial task for building AI systems capable of effective collaboration with humans. Recognising the limitations of humans, who often make suboptimal decisions due to our own fallibility, researchers from MIT and the University of Washington have pioneered a novel approach to address this challenge.

Humans are not capable of infinite computation; hence, they cannot deliberate endlessly to arrive at the perfect solution to a problem. This inherent limitation manifests in what is known as suboptimal decision-making. To model this effectively, the research team developed a method that not only recognises these computational constraints but can infer them from limited observational data.

Reproduced from the paper showing the way the model moves towards the blu and orange star given certain starting conditions. This is how the allowed the agent to explore decision making

The groundbreaking model introduced by the researchers automatically infers an agent’s computational constraints—whether human or machine—by analysing just a few traces of their previous actions. This inference leads to the establishment of an “inference budget,” a predictive tool that can be utilised to anticipate an agent’s future decisions.

At the core of this innovative technique is the concept of modelling behaviour not as a series of random or noise-driven decisions, as many previous approaches have attempted, but as a structured process influenced by the depth and duration of planning an agent can afford given their computational constraints. Inspired by behavioural studies in chess, where the depth of planning varies with the complexity of the situation and the player’s expertise, the researchers have crafted a framework that gauges an agent’s decision-making process by aligning it with a pre-determined computational model running parallel to the agent’s actions.

This method was thoroughly tested in various scenarios, including inferring navigational goals from previously taken routes and predicting moves in chess matches. Remarkably, the researchers’ model not only matched but often outperformed existing methods used in these types of decision-making analyses.

The implications of this research are interesting. By understanding and predicting human behaviour more accurately, AI systems can be designed to interact more effectively with their human counterparts. This could lead to AI agents capable of preempting human errors or adapting to human limitations in real-time, thereby enhancing the collaborative experience.

According to Athul Paul Jacob, the lead author of the study, this modeling technique represents a significant advancement in our ability to make AI truly helpful to humans. “If we know that a human is about to make a mistake, having seen how they have behaved before, the AI agent could step in and offer a better way to do it.” Jacob explains.

Scheduled for presentation at the International Conference on Learning Representations (ICLR 2024) in Vienna, Austria, this research not only advances our understanding of human behaviour in computational terms but also sets a new standard for how we can design AI systems that effectively integrate into human-driven processes.

Looking ahead, the team aims to extend their modelling approach to other domains such as reinforcement learning, commonly used in robotics, underscoring their commitment to enhancing AI-human collaborations across various fields. This work is not just about building smarter machines, but about fostering a synergy that could redefine the boundaries of human-machine interaction.

You can read the full paper “Modeling Boundedly Rational Agents with Latent Inference Budgets” on arXiv.

Staff Writer

Our in-house science writing team has prepared this content specifically for Lab Horizons

Leave a Reply

Your email address will not be published. Required fields are marked *