Led by CELEST board member Jeremy Wolfe
We think of our hunting and gathering stage as confined to dim prehistory but, in fact, while the object of our foraging may have changed over the millennia, foragers we remain. Sometimes, we still search for food – more often at the market than in the field. We gather information. We forage for social cues. We search for signs about the correct path. We do not perform or abandon these tasks at random. We exploit the regularities of the environment to look for things where they are more likely to be and to stop searching when the likelihood is too low. Our goal is to put scientific substance behind assertions of this sort, modeling the behaviors to a level of specificity that will allow us to understand humans and program artificial agents. Thus, the research in this CP is guided by the following question:
How do we learn to forage in a dynamically changing world in order to collect information and resources needed to pursue our goals?
As an exemplar problem, imagine that there has been an industrial accident. Perhaps, as happened in Bangladesh in May 2013, a crowded building has collapsed. The site is unstable and there are many people missing. Some of these have died but others may be alive. A robot, either autonomous or remotely guided, is sent in to examine the site and tell rescue and repair crews where to go. It needs to look for multiple targets over an extended area under time pressure. Having reached a specific location, it is desirable to report all significant information. Some results, such as finding an injured victim, will be more valued than others, such as finding a broken fuel line. However, this could be the only chance to detect the fuel leak. Thus there is a conflict between the desire to acquire high-value items and the desire to be exhaustive. Further, both time and accuracy are important. Delay in finding the wounded could be fatal. False alarms will deflect resources from the real problems.
The world of the collapsed building would be a rule-governed world but those rules would not be the rules pre-loaded into a robot nor would they be the ‘priors’ held by a human operator. For example, expectations about where certain objects would be in an office or factory setting would need to be modified. Effective operation would involve learning how the world has changed.
The task, sketched out above, requires coordination of visual attention, cognition, navigation, and memory in a system capable of learning the regularities of the world. The robot (or its human operator) requires sophisticated abilities to locate and identify the targets of its search. It also needs the ability to decide when to leave the current area and deploy its search capabilities elsewhere. Each of these components has been the object of study in the labs of CELEST researchers associated with the FLAME CP. The goal of the FLAME project is to coordinate our CELEST research efforts. We will create a Virtual Environment (VE) that allows an agent to forage in a navigable environment for desired targets. The VE will be a compromise between the complexity of the real world and the relatively impoverished world of experimental search tasks that each last for only a second or two. Critically, the VE will be a parameterizable world in which rules can be specified and changed with precision.
Each of the labs associated with FLAME is committed to aligning its CELEST-funded research with this unifying FLAME project. FLAME researchers will integrate experimental, modeling, and technological approaches consistent with CELEST’s mission. The development and testing of a computational model capable of performing tasks similar to those studied in the lab and making direct comparisons to experimental results will serve to validate the application of CELEST’s modeling techniques to solve real world problems. As work evolves, a natural extension of this approach is likely to include testing the models in robots.
FLAME research will build upon knowledge from CELEST’s four scientific initiatives. Switching of rules or foraging behaviors is possible because of flexibility, or neural plasticity, in attentional systems. Dynamic coding of information will allow foragers the opportunity to constantly update rules or navigate through changing environments. Functional connections between relevant brain regions will guide model development. Processing bottlenecks in visual attention will guide performance on foraging tasks.