Comments have closed. View all of the finalist entries below.

From Thinking to Inventing

What is the compelling question or challenge?

Alan Turing's question, "Can machines think?" led to significant advances in computing. We propose ushering in a new phase by asking a different question: "what can machines invent and how?"

What do we know now about this Big Idea and what are the key research questions we need to address?

When stuck on a real-world problem, our brains continually restructure how we think about it, generate hypotheses, suggest experiments we can conduct in our environment, make interesting observations, and propose solution strategies that we can execute. While simple to state, this capability is quite complex and remarkable, and as the authors of this submission have shown, it is mathematically intractable [REF1]. Currently, we do not know exactly how humans improvise this way, and we certainly do not have machines that can display similar resourcefulness and resilience.

Cognitive psychology research on the topic is nearly a century old and has attempted to study creative problem solving by observing humans solve insight puzzles. The results have been promising, suggesting that we use various heuristics and techniques in solving these puzzles. Work in cognitive neuroscience has identified certain neural systems that are activated, accompanied by new findings into the role of emotion and mental state that aid effective problem-solving. These results, in turn, have provided a systematic account for uncovering the secrets to insight problem-solving. Nevertheless, these advances have been unable to inform research in artificial intelligence (AI) because the results have been difficult to generalize and operationalize. Without a unified mathematical language to compare one problem with another, or to make the heuristics more concrete and implementable, the results have been largely relegated to a few select, hand-crafted insight problems within the annals of psychology.

Meanwhile, AI research has seen substantial growth and extensive fragmentation into numerous subtopics such as machine learning, automated planning, knowledge representation, computer vision, and many more. Machine learning has been at the forefront of many of AI's recent achievements including victories over humans in intellectual games like Chess and GO. Using Deep Neural Networks (DNNs) these state-of-the-art systems have shown tremendous promise in numerous fields (biology, physics, astronomy, geosciences etc.) However, DNNs have not been successful in handling novel scenarios that require the application of commonsense knowledge in creative ways. This is because, at their core, these systems are fundamentally about finding generalizable patterns in data through mathematical optimization techniques. In many real-world problems, there may be no prior pattern with which to rely on, or worse, there may be no predefined objective function to optimize.

We pose the following research questions:

A. In Theoretical Computer Science

1. How difficult are these problems? Is the Turing Computer a sufficiently good model to capture these computations?

2. Are there complexity classes of these problems, thereby suggesting that one problem might be reducible to another problem in its same class?

3. Are there certain characteristics of creative problems that immediately suggest likely unsolvability?

B. In Cognitive Sciences (Cognitive Psychology, Neuroscience, Computational Modeling)

1. What are the mental representations of creative problems?

2. If the agent can solve a problem using what's available in their environment, then how does an agent select what objects in the environment seem promising for solving their problem?

3. How do the heuristics (like constraint relaxation and chunk decomposition) developed in the insight problem-solving literature generalize across different types of creative problems?

C. In Artificial Intelligence

1. How can an agent decide when the problem is truly unsolvable versus when it should try harder for creative solutions?

2. How can an agent choose what explorations to perform in its environment in order to extract hidden properties useful to problem-solving?

3. How can an agent learn and improve its problem-solving capabilities?

4. How can we evaluate an agent's creative problem-solving capabilities?

Show More


National Science Foundation, 2415 Eisenhower Avenue, Alexandria VA 22314, USA

Tel: (703) 292-5111, FIRS: (800) 877-8339 | TDD: (800) 281-8749