AI Predicts Experiment Outcomes Using Game Theory

Scientists are exploring new avenues to reconcile quantum theory with classical physics by challenging fundamental assumptions about free choice in experiments. Florian Pauschitz from ETH Zurich, Ben Moseley from Imperial College, and Ghislain Fourny from ETH Zurich, detail a novel framework leveraging artificial intelligence to predict outcomes in complex quantum scenarios. Their research, conducted in collaboration between ETH Zurich and Imperial College London, interprets these experiments as deterministic games, modelling the universe as an agent seeking to minimise action. This approach, which abandons traditional assumptions of free choice in favour of ‘contingent free choice’, learns hidden variables using neural networks and represents a significant step towards potentially discovering a local hidden-variable theory extending current quantum mechanics, as demonstrated through a proof-of-concept application to the EPR 2-2-2 experiment. Imagine predicting every move in a complex chess game before it happens. This effort applies artificial intelligence to anticipate outcomes in quantum experiments, challenging long-held assumptions about randomness. By modelling these scenarios as games of strategy, it seeks a deeper, predictable logic underlying quantum behaviour. Scientists are developing a new framework to predict the outcomes of complex quantum experiments, potentially reshaping our understanding of local realism. For decades, interpretations of quantum mechanics have often necessitated abandoning the idea that objects possess definite properties independent of measurement, a concept known as local realism. Scientists are now exploring a different path, challenging a core assumption underpinning many impossibility theorems: the notion of free choice by observers. Instead of dismissing local realism, this effort proposes a compatibilistic version of free choice, termed contingent free choice. Integrates it into a deterministic model. A team has successfully applied artificial intelligence to learn the underlying rules governing these deterministic experiments — by framing quantum scenarios as games between observers and the universe, treated as an agent seeking to minimise action. They’ve created a system where predictions can be made without relying on the traditional assumption of free choice. This approach moves beyond reproducing quantum results. It aims to model the process itself, offering a potential pathway towards a more complete and locally realistic quantum theory — by constructing this model requires determining the ‘rewards’ within the game, representing the hidden variables that influence outcomes. Rather than seeking an analytical solution, The team employed a neural network to learn these reward functions, training it to match predicted outcomes with those dictated by the extended Born rule. A central tenet of quantum mechanics. With the EPR 2-2-2 experiment, a foundational test of non-locality, as a starting point, they demonstrated the feasibility of this AI-assisted approach. The framework’s success isn’t merely about replicating quantum phenomena, but about achieving this replication within a deterministic system that violates Bell inequalities, a key indicator of non-local realism. By minimising the Kullback-Leibler divergence, a measure of the difference between predicted and observed probability distributions, the AI learned reward functions that closely align with quantum statistics, suggesting a promising direction for future research. The system learns rewards by matching observed outcome statistics with those predicted by quantum theory after running millions of deterministic simulations with random hidden variables. Framework predictions align with EPR 2-2-2 experiment outcomes via learned reward functions Once initial simulations were completed, the framework successfully predicted outcomes of the EPR 2-2-2 experiment by closely matching frequency histograms to those obtained via deterministic runs of the game. Specifically, the Kullback-Leibler divergence was minimised, indicating a strong alignment between predicted and observed results. Although a perfect match was not achieved. Discrepancies stemmed from a finite number of simulated runs and a limited learning duration with real parameter values. Here, the learned reward functions, containing a hidden variable, provided a promising foundation for a local hidden-variable theory. Through employing a parameterised reward ansatz inspired by Bell’s original work, the experimental space was reduced to a single dimension. Meanwhile, the offset between measurement basis choices. It’s adaptability extends beyond such simplified scenarios. Here, the project demonstrates the potential to address more complex experiments lacking suitable initial assumptions, utilising more learnable parameters and alternative learning methods like symbolic regression. It’s ability to model the 2-2-2 EPR experiment serves as a proof-of-concept for a toy local-realist model within non-Nashian quantum theory. At a fundamental level, The project suggests a pathway beyond traditional assumptions of free choice — replacing them with a compatibilistic version termed contingent free choice. Through analysis of the learned rewards across a broader range of quantum experiments, beyond the initial 2-2-2 case, and will be key to inferring the underlying local hidden-variable theory. Through learning the reward functions, the framework offers a solid basis for further expansion towards a complete quantum theory. Reward function learning via Kullback-Leibler divergence and deterministic game resolution At the same time, a neural network served as the core component in learning reward functions within a game-theoretic framework designed to model complex experiments — this effort interprets experiments, such as those examining EPR correlations, as games played between an observer and the universe. Conceptualising the universe as an agent striving to minimise action, and by learning proceeded by training the neural network to approximate these reward functions. This embody a hidden variable representing the universe’s ‘strategy’. Establishing the network’s performance necessitated a specific cost function. The Kullback-Leibler divergence quantified the difference between frequency histograms generated from numerous deterministic game runs and the predictions derived from the extended Born rule. Providing a measure of learning accuracy. Once trained, the network effectively predicted outcomes. For the simulation of experimental runs and comparison with established quantum mechanical predictions. The experimental setup for this learning process centred on the EPR 2-2-2 experiment, chosen as a proof-of-concept for the broader framework. By simulating this scenario, researchers aimed to demonstrate the viability of the non-Nashian approach to modelling local hidden-variable theories. For this initial implementation, a parameterised reward ansatz, inspired by Bell’s original work, reduced the experimental space to a single dimension. The offset between measurement basis choices. Beyond simply achieving predictive accuracy, the methodology prioritised adaptability. Unlike approaches reliant on pre-defined ansatze, this AI-assisted framework is intended to be generic — capable of tackling more complex experiments with increased learnable parameters and alternative learning strategies, such as symbolic regression. This approach lies the potential to move beyond the 2-2-2 experiment and explore a wider range of quantum scenarios, and in the end aiming to infer the underlying local hidden-variable theory. Quantum mechanics reconciled with local realism through artificial intelligence and game theory Once a philosophical dead-end, the search for local hidden-variable theories now benefits from an unexpected ally: artificial intelligence. For decades, physicists have grappled with the implications of quantum mechanics, specifically the non-local correlations that seem to defy classical intuition. Attempts to explain these phenomena with hidden variables repeatedly ran into mathematical roadblocks, most famously Bell’s theorem. This new work sidesteps those theorems by relaxing a key assumption: free choice — it proposes a “contingent free choice”, where the universe, modelled as an economic agent, operates according to predictable strategies. Framing experiments as games between observers and the universe is a striking departure, and by learning the ‘reward functions’ governing this game using neural networks. Researchers have created a model that successfully predicts outcomes within a simplified quantum scenario. Rather than seeking to disprove quantum mechanics, this approach aims to build a more complete picture. One that might eventually reconcile quantum behaviour with a classical worldview. The current demonstration relies on a highly specific experiment and a deterministic framework, limiting its immediate applicability. The true power of this effort lies not in the immediate results, but in the methodology. By shifting the focus from proving or disproving quantum mechanics to modelling the underlying dynamics, it opens up new avenues for exploration. Unlike previous attempts, this AI-assisted framework offers a way to systematically search for hidden variables and test their predictive power. By scaling this approach to more complex experiments remains a significant challenge. The field might see a proliferation of similar game-theoretic models, each attempting to capture different aspects of quantum behaviour. The broader effort could benefit from advances in AI algorithms capable of handling increasingly complex game structures. A successful model could not only provide a deeper understanding of quantum mechanics. But also have implications for fields like cryptography and machine learning, where secure and efficient information processing is central. 👉 More information 🗞 Extending quantum theory with AI-assisted deterministic game theory 🧠 ArXiv: https://arxiv.org/abs/2602.17213 Tags: