Struggling to Find The Perfect Gift For The Quantum AI on Your List? Researchers Say Quantum-Classical AI Can Now Play Atari’s Pong And Breakout

Insider Brief

• Researchers from Technische Universität Wien and Freie Universität Berlin developed a hybrid quantum-classical AI that successfully played Pong and Breakout, demonstrating potential for quantum reinforcement learning.
• The hybrid model matched classical performance in Pong and achieved 84% of the classical model’s score in Breakout, reducing the gap to 13% with optimized parameters.
• While the study found no quantum advantage, it highlights how hybrid systems could combine classical and quantum methods to tackle high-dimensional tasks.

In 1975, just in time for the Christmas season, Atari released its home version of Pong, which became a holiday sensation. Humans quickly mastered this game, not, as it turns out, a certain 9-year old — don’t ask me how I know this.

However, nearly 50 years later, a quantum-classical artificial intelligence (AI) model is taking its turn at playing — and likely performing just as good as certain human players did in the mid-’70

In a new study posted on the preprint server arXiv, a team of researchers from the Vienna University of Technology and the Free University of Berlin report their hybrid quantum-classical learning model has successfully played Atari games, like Pong and Breakout. The model achieved competitive scores in both of the games, according to the study.

The team adds that this experiment wasn’t just a chance to get quantum AI in the arcade, it also demonstrates the potential for quantum reinforcement learning in high-dimensional tasks.

The research, conducted by scientists from Technische Universität Wien and Freie Universität Berlin, explores the use of parameterized quantum circuits (PQCs) — quantum programs with adjustable settings — in reinforcement learning, a branch of machine learning where agents learn to make decisions by interacting with an environment. In this study, the researchers integrated PQCs, which process information based on the principles of quantum mechanics, into a hybrid system with classical neural networks to evaluate their performance on tasks typically dominated by deep learning algorithms.

How much better was this quantum-classical AI? In Pong, the model achieved scores equivalent to the classical model, with both models reaching a mean reward of 20. In Breakout, the hybrid model achieved a mean reward of approximately 84 after 2 million environment steps, compared to the classical model’s 141, reflecting a 41% performance gap. However, when the researchers fine-tuned the settings — optimized hyperparameters — the hybrid model’s performance gap reduced to 13%, the researchers report.

No Quantum Advantage “Breakout” Just Yet

The hybrid model matched the performance of a classical reference model in Pong and came close in Breakout, narrowing a typical performance gap observed in earlier studies.

The researchers write in the study: “Our findings suggest that with proper tuning, hybrid agents can closely match the performance of classical agents subjected to the same constraints in the latent space, highlighting the importance of fair benchmarking. The results presented contribute to our understanding of the interplay between quantum and classical components in hybrid models.”

These results suggest hybrid quantum-classical systems could play a role in computationally intensive tasks that require combining classical pre-processing with quantum processing. As quantum computing remains in its noisy, intermediate-scale (NISQ) era, this study offers insights into quantum’s potential as a supplement to traditional machine learning approaches rather than a replacement.

However, the study also clarified that the model did not exhibit a “quantum advantage,” meaning it did not outperform classical methods in this particular application. Instead, it serves as an important step toward understanding how quantum and classical components can complement one another in practical machine learning tasks.

Methods: Tri-layer Architecture

The researchers designed a hybrid architecture incorporating three main components: classical convolutional layers for dimensionality reduction, a PQC for quantum processing and classical fully connected layers for post-processing. This architecture aimed to overcome a significant limitation of current quantum systems—their inability to directly encode large, high-dimensional datasets due to hardware constraints.

In the experiments, the hybrid model processed simplified inputs from Atari game environments. For example, in Pong, the system learned to predict optimal actions based on the position of game elements. Classical convolutional layers extracted features from the high-dimensional observations, while the PQC applied quantum gates to encode and manipulate these features.

Training involved reinforcement learning techniques, where the system optimized its “policy” by iteratively improving its predictions of future rewards based on past decisions. The researchers also tested hyperparameter adjustments — such as reward scaling and learning rate changes — to fine-tune the interaction between classical and quantum components, improving the system’s learning performance.

Limitations and Challenges

The study offers a few limitations and challenges. First, the hybrid model did not achieve significant performance improvements over its classical counterpart, a crucial benchmark for demonstrating quantum advantage. The study attributes this to the nature of Atari games, which are computationally manageable with classical methods.

The quantum component of the model operates on simulated hardware rather than actual quantum devices. Simulations do not capture the noise and error rates present in real-world quantum systems, which could affect performance. Future work would be needed to assess the robustness of the hybrid model under realistic conditions.

Another limitation lies in scalability. Expanding the latent feature space — the intermediate representation processed by the PQC — would require more qubits and deeper quantum circuits, leading to potential issues such as the “barren plateau” phenomenon, where optimization becomes infeasible due to vanishing gradients.

How Can Quantum AI Level Up?

So, how long before quantum-classical AI will be able to spend their holidays racking up massive scores on Pong and Breakout — and, who knows, maybe eventually Frogger? The researchers indicate it will take some work and outlined several avenues for further exploration. One promising direction is to test hybrid models on quantum hardware to evaluate their performance in noisy environments. Another is to apply the models to tasks where quantum computing is expected to offer an advantage, such as quantum chemistry or combinatorial optimization.

While Atari games are a useful benchmark, they may not fully showcase the strengths of quantum-enhanced models, the study suggests. Identifying problem domains where quantum features — such as superposition and entanglement — provide a computational edge could open new possibilities for hybrid systems.

The Bigger Picture

Beyond the nostalgia and video game puns. this research contributes to the growing field of quantum machine learning, where scientists aim to leverage quantum computing for tasks ranging from data analysis to natural language processing. Hybrid quantum-classical models are particularly appealing in the current era of quantum development, as they combine the strengths of mature classical algorithms with the emerging capabilities of quantum processors.

For a deeper technical dive, please read the arXiv paper. ArXiv is a pre-print server, meaning the study has not officially been peer-reviewed.

The study was conducted by Dominik Freinberger, Julian Lemmel, and Radu Grosu from the Institut für Technische Informatik at Technische Universität Wien in Austria, along with Sofiene Jerbi from the Dahlem Center for Complex Quantum Systems at Freie Universität Berlin in Germany.