In a fascinating study of human social cognition conducted by psychologists Felix Warneken and Michael Tomasello, researchers observed an 18-month-old toddler watching an adult struggle with a cabinet door while carrying books. When the adult accidentally bumped the books against the closed cabinet door and appeared confused, something extraordinary occurred.
The toddler instinctively offered assistance. After interpreting the adult's objective, the child approached the cabinet and opened its doors, enabling the adult to store the books inside. This raises an intriguing question: how could a toddler with minimal life experience accurately understand the adult's intention?
Recently, computer scientists have reframed this question for artificial intelligence: How can we develop machines with similar capabilities?
The essential element for engineering this type of comprehension is arguably what defines our humanity: our imperfections. Just as the toddler could discern the adult's goal from his failed attempt, machines that understand human objectives must account for our erroneous actions and strategies.
In the pursuit of embedding this social intelligence in machines, researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Department of Brain and Cognitive Sciences developed an algorithm capable of deducing goals and plans, even when those plans might be unsuccessful.
This innovative research could potentially enhance various assistive technologies, collaborative or caregiving robots, and digital assistants like Siri and Alexa.
"This capacity to account for mistakes could be essential for constructing machines that reliably infer and act in our best interests," explains Tan Zhi-Xuan, a PhD student in MIT's Department of Electrical Engineering and Computer Science (EECS) and the lead author of a new paper on this research. "Without this capability, AI systems might incorrectly conclude that because we failed to achieve our higher-level objectives, those goals weren't important to us. We've witnessed the consequences when algorithms exploit our impulsive and unplanned social media usage, leading to addiction and polarization. Ideally, future algorithms will recognize our mistakes, poor habits, and irrationalities and help us avoid, rather than reinforce, them."
To develop their model, the team utilized Gen, a novel AI programming platform recently created at MIT, to integrate symbolic AI planning with Bayesian inference. Bayesian inference offers an optimal method to combine uncertain beliefs with new information, and is extensively applied in financial risk assessment, medical diagnostics, and election predictions.
The team's model demonstrated performance 20 to 150 times faster than an existing benchmark method called Bayesian Inverse Reinforcement Learning (BIRL), which learns an agent's objectives, values, or rewards by observing behavior and attempts to compute complete policies or plans in advance. The new model achieved 75% accuracy in inferring goals.
"AI is transitioning away from the 'standard model' where a fixed, known objective is provided to the machine," notes Stuart Russell, the Smith-Zadeh Professor of Engineering at the University of California at Berkeley. "Instead, the machine acknowledges that it doesn't know what we want, which means research on inferring goals and preferences from human behavior becomes central to AI. This paper embraces that objective; specifically, it represents progress toward modeling—and therefore reversing—the actual process by which humans generate behavior from goals and preferences."
How it works
Although significant research exists on inferring agents' goals and desires, much of this work assumes that agents act optimally to achieve their objectives.
However, the team drew particular inspiration from a common human planning approach that's largely sub-optimal: rather than planning everything in advance, humans tend to form partial plans, execute them, and then plan again from that point. While this can lead to errors from insufficient foresight, it also reduces cognitive burden.
For instance, imagine observing a friend preparing food and wanting to help by determining what they're cooking. You might predict the next few steps your friend could take: perhaps preheating the oven, then preparing dough for an apple pie. You then "retain" only the partial plans that remain consistent with your friend's actual actions, and repeat the process by planning just a few steps ahead from there.
Once you've seen your friend make the dough, you can narrow the possibilities to baked goods only, and guess they might slice apples next, or gather pecans for a pie mixture. Eventually, you'll eliminate all plans for dishes your friend couldn't possibly be making, keeping only the feasible options (i.e., pie recipes). Once you're reasonably certain which dish it is, you can offer assistance.
The team's inference algorithm, called "Sequential Inverse Plan Search (SIPS)", follows this sequence to deduce an agent's goals, as it only creates partial plans at each step and eliminates unlikely options early. Since the model only plans a few steps ahead each time, it also considers the possibility that the agent—your friend—might be doing the same. This includes the potential for mistakes due to limited planning, such as not realizing you might need both hands free before opening the refrigerator. By detecting these potential failures in advance, the team hopes the model could enable machines to provide better assistance.
"One of our early insights was that if you want to infer someone's goals, you don't need to think further ahead than they do. We realized this could be used not just to accelerate goal inference, but also to deduce intended goals from actions that are too shortsighted to succeed, leading us to shift from scaling up algorithms to exploring ways to address more fundamental limitations of current AI systems," explains Vikash Mansinghka, a principal research scientist at MIT and one of Tan Zhi-Xuan's co-advisors, along with Joshua Tenenbaum, MIT professor in brain and cognitive sciences. "This is part of our larger moonshot—to reverse-engineer 18-month-old human common sense."
The research builds conceptually on earlier cognitive models from Tenenbaum's group, demonstrating how simpler inferences that children and even 10-month-old infants make about others' goals can be quantitatively modeled as a form of Bayesian inverse planning.
While to date the researchers have only explored inference in relatively small planning problems over fixed sets of goals, through future work they plan to investigate richer hierarchies of human goals and plans. By encoding or learning these hierarchies, machines might be able to infer a much broader range of goals, as well as the deeper purposes they serve.
"Though this work represents only a small initial step, my hope is that this research will establish some of the philosophical and conceptual foundations necessary to build machines that truly understand human goals, plans, and values," says Xuan. "This basic approach of modeling humans as imperfect reasoners seems very promising. It now allows us to infer when plans are mistaken, and perhaps it will eventually enable us to infer when people hold mistaken beliefs, assumptions, and guiding principles as well."
Zhi-Xuan, Mansinghka, and Tenenbaum authored the paper alongside EECS graduate student Jordyn Mann and PhD student Tom Silver. They virtually presented their work last week at the Conference on Neural Information Processing Systems (NeurIPS 2020).
This work was funded, in part, by the DARPA Machine Common Sense program, the Aphorism Foundation, the Siegel Family Foundation, the MIT-IBM Watson AI Lab, and the Intel Probabilistic Computing Center. Tom Silver is supported by an NSF Graduate Research Fellowship.
