Synthesizing intelligent behavior

- Research Unit 1 -

This research unit implements our synthetic approach to intelligence research. Projects synthesize intelligent behaviors based on the proposed loop linking analytic and synthetic disciplines of intelligence research.

We conduct research in three subunits, each pertaining to a specific type of intelligence and a corresponding intelligent example behavior. Example behaviors are chosen to capture fundamental aspects of individual intelligence, social intelligence, and collective intelligence. For the scope of this cluster, individual intelligence is attributed to a single agent behaving in its environment (e.g. assembling a shelf without a manual). Social intelligence is exhibited when several agents that exhibit individual intelligence interact with each other (e.g. hiding an object from another person). Collective intelligence is exhibited by many agents, but in contrast to social intelligence, the resulting behavior cannot be attributed to any of the individuals but only to their collective (e.g. a swarm of fish escaping a predator).

Example behaviors serve as a touchstone for intelligence: We chose the three example behaviors so that an agent enacting them would be considered intelligent. The analysis and synthesis of the example behaviors then serves as a touchstone of our understanding of intelligence. To avoid the historical mistakes of AI, we also require target behaviors to be performed in domains in which AI has not traditionally been successful. We therefore selected three example behaviors that must be performed in the real world:

 

Behavior 1 - Individual intelligence: Escaping from an escape room

You are locked in a room. To escape you must solve a number of puzzles that will unlock the door. Everything you need for your escape is inside the room. But can you find the clues and put them together? This is the setting in the Escape Room, a real-world adventure game that has attained significant popularity in recent years.

To escape from an escape room, a person must search the room for clues, such as keys, levers, buttons, hidden doors, etc. But there might also be objects hidden in cupboards that reveal part of the escape route. For example, in a real escape room you might find jars of different weights, each with a number on it, and a balance scale. Ordering the jars by weight reveals the code to a number lock that then reveals the next clue.

Escaping from an escape room demonstrates intelligence, according to our preliminary definition of intelligence. The behavior is goal-directed (escape). It requires adaptation at many levels (the clues will be different each time and they will be interdependent in novel ways). The behavior must be performed in a cost-effective manner (the combinatorics of all clues in an escape room are otherwise prohibitive). And finally, the behavior exhibits generality, as the types of puzzles that have to be solved can be quite different within the same escape room (or even in different escape rooms).

Following our synthetic approach to intelligence research, we will study human behavior in escape room settings and use the insights gained to synthesize an escape room robot. We will then use the synthetic artifacts as tools to contribute to the analytic disciplines, closing the loop between analytic and synthetic disciplines.

 

Behavior 2 - Social intelligence: Learning in social interaction

Imagine playing a jigsaw puzzle game together with your child. You are both shifting the jigsaw puzzle pieces around on the floor in order to successfully solve the game together. Your child is struggling to place the next piece of the puzzle. You would like to help, but you do not simply want to show the correct location of each piece. Instead, you would like to communicate to your child a helpful, more general rule - something that will also help with another puzzle. You could, for example, show that starting with the corner and edge pieces is a good idea, or that light blue pieces are probably part of the sky and should be placed in the upper part of the puzzle. By doing so you are passing on your own knowledge and experience on to your child through social interaction.

The ability to transfer knowledge among conspecifics is an important hallmark of human and animal intelligence. Humans are far better at communicating, accumulating, and extending knowledge in the context of a social group than any other species - although animals such as mice or fish do also transfer knowledge socially. If this transfer of knowledge is more efficient than the acquisition of the same knowledge by an individual alone, then it represents an important contribution to the intelligent behavior of an agent. To fully understand social intelligence and to replicate the social transfer of knowledge in technological artifacts, we must therefore provide effective means of knowledge transfer between agents. Related abilities go beyond the abilities of individual intelligence discussed in the previous section (Behavior 1); social intelligence can be viewed as a behavioral level that is distinct from the aspects of individual intelligence covered by the escape room scenario.

The effectiveness of social knowledge transfer in biological agents depends on many factors. In humans, for example, transfer occurs via different cues and across multiple sensory modalities, including verbal cues, pointing gestures, facial expression, or observable actions (e.g. moving a puzzle piece to a particular location). A teacher has to select among these possibilities the actions that maximize the efficiency of knowledge transfer to the learner, who has to detect and interpret them correctly. The social intelligence of an agent can then be understood as the degree of efficiency of social knowledge transfer. This includes the ability to communicate information about activities, solutions, and subgoals. Social interactions support the learning of new strategies and increase learning speed.

To teach effectively, the teacher must (1) possess an understanding of the task and the learner, (2) select actions tailored to the knowledge available to the learner and (3) be able to communicate effectively in a social setting.

In most social settings, the roles of teacher and learner are not fixed, but alternate between the agents during turn-taking behavior. In the analytic disciplines, we will study these aspects in humans, mice, and fish. In accordance with our research strategy, resulting insights will be encoded in synthetic artifacts to enable the effective knowledge transfer among synthetic agents but also between synthetic and biological agents.

Jigsaw puzzle task: To increase the complexity of the social interaction beyond the scenario of a child’s puzzle, where knowledge pertains to the shape of the pieces and the partial image on them, we use the general idea of a game to encode a more complex task. One such example is Sudoku, where the puzzle pieces represent numbers that have to be brought into a specific arrangement. The rules of this arrangement and strategies for solving a Sudoku puzzle would then have to be transferred between agents in a social learning task. We can vary the complexity of the task on two different axes: By increasing the complexity of the puzzle, we can scale the example behavior to our increasing understanding of knowledge transfer in a social setting. By increasing the complexity of the social component in the game, e.g. by playing an interactive game such as Memory, Uno, or Settlers of Catan, we can scale the amount of social understanding required.

 

.

Behavior 3 - Collective intelligence: Cooperative shepherding

Consider a group of heterogeneous artificial agents, “shepherds”, each with restricted sensory and cognitive abilities, placed in a novel and complex environment containing various obstacles interfering with their locomotion and perception. The agents have to form a collective and coordinate to find, collect, and guide a large number of artificial “sheep” agents, distributed in the environment. The “sheep” interact with each other based on behavioral rules unknown to the shepherds. Furthermore, individual shepherds have no initial knowledge of how many other shepherds are present and what their sensory and cognitive abilities are. Given this scenario, the self-organized shepherd collective must manipulate the sheep according to a pre-set objective under temporal constraints, for example: “Move all the sheep to a distinct target region along a specific target trajectory”, or “Collect at least N sheep and separate them into M groups of the same size.”

This behavior exemplifies a fascinating phenomenon observable all across nature: collectives, starting from social insects to human societies, exhibiting adaptable, cost-effective and goal-directed collective behavior. According to our working definition of intelligence, they exhibit collective intelligence. Individuals within the collective may not even be aware of the overall goal. Each might just follow a set of simple behavioral rules, and yet the collective as a whole is capable of solving problems beyond the reach of single individuals. A beautiful example of such behavior is the construction of “living bridges” by army ants, collectively assembled from their own bodies to span gaps obstructing the colonies path (see e.g. [1,2]). Here, intelligence is not only functionally non-decomposable, but it is also distributed across the collective. It can not be simply reduced to intelligent behaviors of individuals (Behavior 1) or easily scaled up from few interacting intelligent agents (Behavior 2).

Following [3], we specify the following defining characteristics of collective intelligence:

  1. There is no centralized control, no single leader;
  2. Individuals may independently acquire information, but no agent has direct access to the global state of the system;
  3. Information is combined and processed within the collective through social interactions;
  4. A problem is solved using a strategy that cannot be implemented by a single individual.

These criteria also apply to our shepherding example, which in turn is related to many problems in analytic and synthetic sciences (see e.g. [4,5,6]). Synthesizing this complex collective behavior will boost our understanding of collective intelligence across different systems including groups of humans working together towards a solution of a complex problem in more abstract settings. Furthermore, the above shepherding scenario offers also many options for future extensions, e.g. the introduction of adaptation of the sheep to the shepherds, eventually leading to a co-evolution of two intelligent collectives.

The situation of collective intelligence research is similar to intelligence research in general. Different disciplines study aspects of collective intelligence in isolation, e.g. collective perception and movement coordination in biology and swarm robotics (see e.g. [7,8]), or the question on how to quantify collective intelligence in psychology [9]. Yet these research efforts remain fragmented, and we are as far from understanding collective intelligence as a complex phenomenon as in individual intelligence research. By combining research on animal groups and synthetic swarms, we aim at uncovering general principles of collective intelligence as a distributed phenomenon which cannot be trivially traced back to individual intelligence. In particular, we will explore not only how the collective complex behavior decomposes into simpler behavioral components, but also how these different components must be distributed across the collective in order to facilitate collectively intelligent behavior.

Understanding the designated collective behavior requires the integration of various fields, including behavioral biology, neuroscience, and cognitive science. The complexity of the integrated shepherding task makes it virtually impossible to study it in an analytic setting alone (biology or psychology) while controlling all relevant aspects. The successful synthesis of the example behavior requires tight integration of analytic and synthetic disciplines for understanding crucial aspects of collective behavior such as coordination, information exchange and decision making - and their interplay. By providing computational and morphological constraints, a synthetic implementation offers a principled approach to exploring the components that are necessary and sufficient for the collective intelligence inherent to this behavior.

Last but not least, the synthesis of self-organized, intelligent collectives raises a whole set of ethical issues. These include questions of the controllability of such distributed systems and possible civilian applications e.g. in preventing crowd disasters [10] but also the possibility of military applications. Thus, despite the fundamental nature of the proposed research, it will, from the beginning, be accompanied by a critical ethical reflection as part of the responsible research and innovation strategy within SCIoI (see Research Unit 4 - Sociological and philosophical framing).

Research activities will go beyond the example behaviors

The three example behaviors provide a unifying objective and a force towards cohesion for the disciplines of intelligence research. However, our research will not be confined by these example behaviors. It is therefore important to note that in the following three sections we will describe research components which contribute to the three example behaviors but will also investigate research questions that are more fundamental and go beyond the specific behavior. These activities, too, will synthesize intelligent behavior following our synthetic approach to intelligence research.

Example behaviors accomplish cohesion

We acknowledge that the starting point of our efforts is today’s fragmented but nevertheless foundational intelligence research. We will therefore continue to rely on the established functional subdivisions (e.g. perception, action, reasoning, etc.) to structure the description of our activities. This will also allow a more natural presentation of related work. However, our focus on example behaviors enforces the merging of all functional components into a single, integrated synthetic artifact. Behaviors are indeed orthogonal to the functional decomposition established in intelligence research. They require the close integration of these components. By exhibiting a behavior in a single synthetic artifact (one artifact for each of the three example behaviors), we will gain an understanding of the interactions between components and therefore of the non-decomposability of intelligence.

Our research will explicitly consider and reflect dependencies among research components, in an effort to achieve cohesion among them and to understand the non-decomposability of intelligence.

References

[1] Self-assemblages in insect societies - Anderson, Carl and Theraulaz, Guy and Deneubourg, J-L. Insectes sociaux Vol. 49-2, pp. 99-110, 2002 Springer

[2] Army ants dynamically adjust living bridges in response to a cost-benefit trade-off - Reid, Chris R and Lutz, Matthew J and Powell, Scott and Kao, Albert B and Couzin, Iain D and Garnier, Simon. Proceedings of the National Academy of Sciences Vol. 112-49, pp. 15113-15118, 2015 National Academy of Sciences

[3] Swarm intelligence in animals and humans - Krause, Jens and Ruxton, Graeme D and Krause, Stefan. Trends in Ecology & Evolution Vol. 25-1, pp. 28-34, 2010 Elsevier

[4] Shepherding behaviors - Lien, Jyh-Ming and Bayazit, O Burchan and Sowell, Ross T and Rodriguez, Samuel and Amato, Nancy M. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) Vol. 4, pp. 4159-4164, 2004

[5] Solving the shepherding problem: Heuristics for herding autonomous, interacting agents - Strömbom, Daniel and Mann, Richard P and Wilson, Alan M and Hailes, Stephen and Morton, A Jennifer and Sumpter, David J T and King, Andrew J.Journal of the Royal Society Interface Vol. 11-100, pp. 20140719, 2004 The Royal Society

[6] Autonomous Shepherding Behaviors of Multiple Target Steering Robots - Lee, Wonki and Kim, DaeEun. Sensors Vol. 17-12, pp. 2729, 2017 Multidisciplinary Digital Publishing Institute

[7] Collective memory and spatial sorting in animal groups - Couzin, Iain D and Krause, Jens and James, Richard and Ruxton, Graeme D and Franks, Nigel R. Journal of Theoretical Biology Vol. 218-1, pp. 1-11, 2002 Elsevier

[8] Swarm robotics: A review from the swarm engineering perspective - Brambilla, Manuele and Ferrante, Eliseo and Birattari, Mauro and Dorigo, Marco. Swarm Intelligence Vol. 7-1, pp. 1-41, 2013 Springer

[9] Evidence for a collective intelligence factor in the performance of human groups - Woolley, Anita Williams and Chabris, Christopher F and Pentland, Alex and Hashmi, Nada and Malone, Thomas W. Science Vol. 330-6004, pp. 686-688, 2010 American Association for the Advancement of Science

[10] Saving human lives: What complexity science and information systems can contribute - Helbing, Dirk and Brockmann, Dirk and Chadefaux, Thomas and Donnay, Karsten and Blanke, Ulf and Woolley-Meza, Olivia and Moussaid, Mehdi and Johansson, Anders and Krause, Jens and Schutte, Sebastian and others. Journal of Statistical Physics Vol. 158-3, pp. 735-781, 2015

We are hiring!

We are looking to hire main office staff (Managing Director, Administrative Assistant and Finance Officer) right now!

View open positions