Few-Shot Design Optimization by Exploiting Auxiliary Information

Many real-world design problems involve optimizing an expensive black-box function $f(x)$, for which Bayesian Optimization is a sample-efficient framework. However, while the basic black-box setting returns a scalar reward, real-world experiments often generate a wealth of useful information. We introduce a new setting where an experiment generates high-dimensional auxiliary information $h(x)$ along with $f(x)$; moreover, a history of relevant, previously-solved tasks is available for accelerating optimization. We develop a novel method based on a neural model which predicts $f(x)$ for unseen designs given a few-shot context containing observations of $h(x)$. We evaluate our method on two challenging domains, robotic hardware design and hyperparameter tuning. On both domains, our method achieves improved few-shot prediction and faster design optimization, outperforming several multi-task optimization methods.

Design problems are ubiquitous across engineering and the natural sciences. For instance, a biologist may want to design a drug that binds as tightly as possible to a pathogen, and a roboticist might want to design a robot arm that can grasp objects delicately and stably. Often, designing good solutions involves running real-world experiments, e.g. a wet-lab experiment to test drug binding. This experiment is a "black-box", where you can put in a design, and you get out a number measuring how well the design performed. The goal is to optimize this metric with as few experiments as possible. A number of AI methods exist that intelligently decide which design to try next, based on the experiments that have come before. However, this basic "black-box" setting, which only returns a single number measuring a design, is highly simplified. Modern scientific or engineering labs have advanced experimental capabilities, which are capable of making several observations about a system. For instance, in robot design, trying out a robot arm may generate a high volume of sensor data (from cameras, or tactile sensors) along with a final performance measure of the design. Therefore, we introduce a new optimization setting, where a trial generates high-dimensional "extra information" along with the number measuring performance. This extra information can be very useful for understanding not just that a design fails, but how exactly the design fails, and could be altered to succeed. We introduce a novel AI method for this setting. This method involves a neural network model, which is trained on a history of design tasks that have already been solved. It learns how to take in a small set of evaluated designs for a task, which includes observations of this 'extra information', and predict which un-evaluated designs might have high reward and should be tried out. After being trained, this model can be applied to a new design task, where it iteratively predicts which design to try next, and tries out that design, repeating until a satisfactory design is found. We apply our method to multiple design problems. One problem involves designing robotic grippers whose shape must be customized to grasp specific objects (e.g. a bottle). Each time the gripper makes contact with the object, it gets `extra' tactile feedback along with a reward for the grasp. We show that our method can grasp new objects that it did not see during training, and can quickly optimize the gripper design for the object after only a few interactions with that object. It finds successful gripper designs for a new object significantly faster than current methods. Thus, our work takes a step towards more capable systems for AI-driven design, which can conduct effective design and discovery in realistic scientific and engineering environments.