Abstract
This paper presents a framework for Cooperative Bayesian Optimization (CBO) where an AI agent and a human user collaborate to optimize an unknown black-box function. Unlike standard Bayesian Optimization where a single agent has full control, CBO involves partitioned action spaces. By modeling the human partner as a "computationally rational" but imperfect agent—characterized by conservatism, explorativeness, and decision noise—the AI can employ strategic planning via a Partially Observable Markov Decision Process (POMDP). Experimental results on synthetic benchmarks demonstrate that a strategic AI, which anticipates and influences human behavior, significantly outperforms greedy approaches that ignore the partner's cognitive biases.
Introduction
In many real-world optimization tasks, such as molecular discovery or collaborative design, an AI system must work alongside a human expert. Often, the action space is split: the AI controls one set of parameters (e.g., temperature), while the human controls another (e.g., concentration). This "Human-AI Teaming" scenario is challenging because humans are not perfect optimizers. They possess limited information-processing capabilities and inherent cognitive biases. If an AI acts "greedily" by only considering its own immediate knowledge, it may fail to guide the team out of local optima if the human partner is biased or conservative in their belief updates.
Methodology
The authors frame the cooperative task as a repeated Bayesian game. To solve this, they implement a Strategic AI agent using a Partially Observable Markov Decision Process (POMDP) framework, solved via Bayes Adaptive Monte Carlo Planning. This allows the AI to perform "zero-shot" planning, simulating future interactions to maximize long-term rewards rather than immediate gains.
A critical component is the User Model. The AI maintains a "theory of mind" by modeling the human as a computationally rational agent with three specific imperfections:
- Conservatism ($\alpha$): Resistance to updating prior beliefs with new evidence.
- Explorativeness ($\beta$): The balance between exploration and exploitation in the user's acquisition function.
- Decision Noise: Stochasticity in the user's final choice, modeled using Thurstone’s law.
The AI estimates these parameters online using Laplace approximation, allowing it to adapt its strategy to the specific behavioral profile of its partner.
Experimental Setup
The framework was evaluated using synthetic experiments on a 3-modal variant of the Himmelblau function. This function provides multiple peaks, testing the agent's ability to avoid local optima. Synthetic user profiles were generated with varying levels of conservatism and exploration. The performance of the Strategic AI was compared against a "Greedy AI" (which ignores the user's internal state) and a "Random AI."
Results and Discussion
The Strategic AI consistently outperformed both Greedy and Random baselines. While the Greedy AI often found local optima quickly, it frequently became stuck because it failed to "nudge" the human toward better regions of the search space. In contrast, the Strategic AI was willing to take seemingly sub-optimal short-term actions to influence the human's beliefs and lead the team to the global maximum.
Strategic planning proved most beneficial when the human partner was highly conservative. Interestingly, while the Strategic AI was superior at optimization, a Random AI was better at reducing overall uncertainty about the function. This highlights a trade-off between strategic exploitation (winning) and pure information gathering (learning).
Conclusion
This work demonstrates the importance of modeling partner behavior in human-AI collaborative optimization. By accounting for human cognitive biases, AI agents can more effectively navigate complex search spaces. Future research will focus on reducing the computational overhead of the planning algorithm to enable real-time interaction and testing the framework with real human participants to validate the computational rationality model in more complex, high-dimensional environments.
Summary updated