Harvey Mudd's virtual gopher study examines how detection of deliberate traps improves survival
All human and animal life involves some level of risk assessment in daily life. Now a group of researchers, led by George Montanez, an assistant professor at Harvey Mudd College in California, has used computer simulations to quantify how the ability to calculate risk can affect survival.
All human and animal life involves some level of risk assessment in daily life. Now a group of researchers, led by George Montanez, an assistant professor at Harvey Mudd College in California, has used computer simulations to quantify how the ability to calculate risk can affect survival.
Specifically the researchers investigated whether a virtual agent's ability to perceive the intention behind a potential risk made a difference in survival. In terms of artificial intelligence (AI), the question is important for how robots will interact with humans in the future and how they will deal with potential danger.
"Our study represents a niche but cutting-edge area of AI research, investigating how equipping artificial agents with a simple theory of mind can benefit them in adversarial situations," Montanez said.
The computer simulation created virtual gophers and "tested whether intention could be perceived through artifacts, namely, structures in an environment that could potentially be traps," Montanez added. Further, the study looked at whether the virtual gopher's "knowledge of whether the structures were intended or accidental could improve survival rates for artificial agents (in our case, artificial gophers)."
The experimental setup
Montanez explained the two-fold risks for the artificial gophers in the experimental setup.
"All the structures in our experiments, whether traps or not, contain food, and the gophers are incentivized to obtain food or starve," he said. "Avoiding too many structures results in starvation but entering too many structures increases the chances of walking into a designed trap."
The results of the computer simulations showed that the gophers who demonstrated intention perception and avoided the designed traps had advantages over those gophers without intention perception.
"Being able to perceive the intended purpose of a structure and make decisions based on that knowledge gave the intention-perception gophers in our experiments significant and measurable advantages, allowing them to outlive their intention-blind counterparts," Montanez said.
How it worked
The computer simulation involved a grid with 12 boxes with food in the center and a system of wires and arrows connected to arrows, which function as "projectile laser beams." The virtual gopher had to navigate the boxes with wires without setting off an arrow beam, "which could potentially hit a gopher in the structure and kill it," depending on the direction of the arrow.
Some of the gophers could distinguish structures that were intended, coherent traps or randomly constructed.
The researchers put the virtual gophers through 50 structures, some with deliberate traps and others with random configurations. They repeated this many times, and "measured how many structures the gophers went through before dying, whether they starved, and whether they successfully avoided actual traps," Montanez said.
A description of the experimental setup, with pictures of the traps and gophers, can be found here.
The researchers found that "those gophers who were able to accurately determine whether the structures were designed with the intent to harm allowed the intention-perception gophers to avoid many lethal traps and, therefore, live longer.
"While they also became less likely to enter structures overall, increasing their chances of starvation, "the tradeoff clearly worked in their favor, with them outlasting gophers without intention perception by a large margin," Montanez said.
Montanez added, "While we initially hypothesized that intention perception would give an advantage to agents, we didn't expect the advantages to be so significant."
The research team
This research was one of three studies on intention perception carried out by Montanez's undergraduate students, and presented at peer-reviewed conferences. It was supported by the National Science Foundation.
The work continues, Montanez said.
"Our long-term goals are to formalize the science of intention perception, building reliable sensors that work in a variety of scenarios, and test what advantages can be gained by having these abilities in our robotic and software agents," he said.
Aside from his own contributions, Montanez stressed, "all this work and its follow-up was done solely by undergraduate students Amani Maina-Kilaas, Cynthia Hom, Anshul Kamath, and Nick Grisanti (Harvey Mudd College), Kevin Ginta (Biola University); Jiayi Zhao (Pomona College); and Cindy Lay (Claremont-McKenna College). More information about our lab and amazing students can be found at https://cs.hmc.edu/AMISTAD."
----
C. Hom et al., "The Gopher's Gambit: Survival Advantages of Artifact-Based Intention Perception." 13th International Conference on Agents and Artificial Intelligence (ICAART 2021), Online, Feb 4–6, 2021.
Alerts Sign-up