| The evolution and the neural mechanisms of cooperation and communication | ||
|
||
|
Social behavior can be found on almost every level of life, ranging from microorganisms to human societies. However, explaining the evolutionary emergence of cooperation, communication, or competition still challenges modern biology. The most common approaches to this problem are based on game-theoretic models. The problem is that these models often assume fixed and limited rules and actions that individual agents can choose from, which excludes the dynamical nature of the mechanisms that underlie the behavior of living systems. So far, there exists a lack of convincing modeling approaches to investigate the emergence of social behavior from a mechanistic and evolutionary perspective. Instead of studying animals, the methodology, we employ, combines several aspects from alternative approaches to study behavior in a rather novel way. Robotic models are considered as individual agents which are controlled by recurrent neural networks representing non-linear dynamical system. The topology and parameters of these networks are evolved following an open-ended evolution approach, that is, individuals are not evaluated on high-level goals or optimized for specific functions. Instead, agents compete for limited resources to enhance their chance of survival. Further, there is no restriction with respect to how individuals interact with their environment or with each other. As our main objective, we aim at a complementary approach for studying not only the evolution, but also the mechanisms of basic forms of communication. For this purpose a robot does not necessarily have to be as complex as a human, not even as complex as a bacterium. The strength of this approach is that it deals with rather simple, yet complete and situated systems, facing similar real world problems as animals do, such as sensory noise or dynamically changing environments. |
||
|
|
||
|
Principles of structural coupling (cf. Maturana and Varela, 1987): Due to structure determined and structure determining interaction of an unity with its environment or another unity each interacting system is a source (and a target) of pertubations with respect to the other. | |
|
The used evolutionary algorithm ENS^3 (cf. Intelligent Dynamics Group the Fraunhofer AIS) optimizes not only the parameter of RNNs but also their topology. Therefore any kind of dynamics can emerge. Furthermore this allows us to evolve small sized networks which are still analytical tractable which enables us to relate the dynamics of the RNNs to the observed behavior. | |
|
|
An example: The small sized RNN evolved for a light seeking task (click on the pic to enlarge it) in simulation exhibits such an robust behavior that it can be easily transferd to real robots without any modifcation (see movie on the right side). |
| Aggressive communication and competitive behavior | ||
| the project | under construction... (coming soon). | |
|
|
||
| related publications |
|
|
| Emergent communication and signal coordination | ||
| the project |
One important aspect of biological evolution is the variety of solutions
developed by nature for similar problems. In this project we investigated the
role of evolutionary variety concerning different neural mechanisms that
underlie seemingly similar behaviors among cooperating agents. Here,
cooperation was based on simple communication forms which are reminiscent of
food or alarm calls among animals. Questions pursued in this project are:
Thus, even though the evolved behaviors are seemingly similar, evolution develops a variety of clearly distinct neural mechanisms which realize these behaviors. And this diversity of mechanisms is in fact important for the flexibility of groups of agents when environmental conditions change because some of them possess an intrinsic robustness to these changes. |
|
|
|
||
| movies |
|
Solitary foraging behavior which evolved when individuals were not able to communicate. |
|
Evolved food calls. Individuals which find a food source guide their conspecifics to the source via acoustic signals (robot turns red). This solution is based on neural oscillators which become activated when a robot discovers a food source. | |
|
A different solution evolved under the same evolutionary conditions. The seemingly similar behavior is, however, mainly driven by sensorimotor noise. Robots also coordinate their signals indirectly by activating the infrared sensor of other individuals on the food spot which, as a consequence cease signaling. | |
|
Increasing the number of interacting individuals posses an interference problem of uncoordinated sound signals. In this solution food calls become synchronized through local interactions. This is based on pulse coupled oscillators. | |
|
|
||
| related publications |
|
|
| It's all about rhythm: Behavior coordination through synchronized communication | ||
| the project |
In this project we wanted to investigated how, given our dynamical (neural)
system approach, can self-organized and coordinated (i.e., synchronized)
behavior among autonomous robots be achieved with minimized communication
efforts?
Therefore, we considered a setup where robots had to collect energy (i.e., foraging behavior) in their environment and had to convey that energy to a base (i.e., homing behavior) where it is stored and constantly consumed. To display one of the both behaviors, either foraging or homing, robots were equipped with a structurally small neural rhythm generator, a resettable oscillator inherent in the robot's neural control which periodically inhibited specific sub-behaviors. To coordinate the behavior among many individuals driven by such an internal rhythm, individuals could communicate with each other via simple acoustic signals, reminiscent of the flashes used by fireflies. Communication among the robots linked what was internal to them, their rhythm, to what was external to them, their foraging and homing behavior. In that sense we went a step further to what Parisi (2004) proposed as "Internal Robotics". We considered not only the interaction of an internal drive with environmental stimuli which externally drives an individual, we also considered how a whole population coordinate the internal drive of each individual through local couplings realized by minimalistic acoustic communication. We showed that synchronizing behaviors among robots based on this rather simple, yet highly efficient, mechanism is not only scalable but also remarkably robust to the spatial range within individuals can interact with each other. |
|
|
|
||
| movies |
|
25 robots forage for three food sources (yellow blobs) and return, with the collected energy, back to their nest (blue circle). About minute 10 individuals start to coordinate their foraging and homing behavior in a group via acoustic communication (red spheres indicate sound signals and their reception range). Behavior coordination by synchronization increases foraging efficiency as indicated by the periodic increase of collected energy within the nest. |
|
The same network applied to 150 robots shows the scalability and robustness of this synchronization process. | |
|
|
||
| related publications |
|
|
| Emergent clustering: From pursuit to aggregation | ||
| the project |
Aggregation is a prevalent pattern found in biological systems. The classical
interpretation from an evolutionary perspective is that aggregation gives a
group of animals advantages regarding mate choice, collective information
sharing, or increased protection from predators. Aggregation can be found on
any scale, from unicellular organisms to whales, from small groups to millions
of individuals. Interestingly, such self-organized group behavior can be
explained by a few, yet simple, local interaction rules which are based on a
balance between positive and negative feedback.
In this project we investigated (i) how individual behavior can be evolved which shows similar attraction and repulsion properties and (ii) which collective phenomena can be observed when many individuals interact with each other. Our experiments revealed that the integration of two competing behaviors, repulsion from obstacles and attraction to sound sources, can by realized by a differently weighted influence of different sensor modalities, that is, by a different impact of the changing parameters of a neural network. Interestingly, both behaviors result from the very same neural dynamics on which the different sensor modalities act on. The neural mechanisms which enabled a single robot to handle difficult environmental conditions are based on hysteresis effects, caused by bi-stable regions within the control system. We were then able to use the very same structurally small control system, actually evolved for an individual behavior, to accomplish complicated aggregation patterns, solely based on simple local interactions. All what was needed was to make every robot in a large group a potential target for all other robots. That is, the robots by themselves realized the link between individual behaviors and a global coherent pattern, such as aggregation, without the need of centralized or hierarchical control. |
|
|
|
||
| movies |
|
Three robots (green) evolved to pursue the red robot which continuously emits an acoustic signal (indicated by the red sphere). |
|
120 robots controlled by the very same simple neural network, but this time all robots emit a sound signal. As a result of the local robot interactions aggregation patterns emerge. | |
|
|
||
| related publications |
|
|
| Tied together: Evolution of cooperation without explicit communication | ||
| the project |
Many biological examples, such as ant societies or schools of fish, exhibit
complex collective behavior patterns while the behavioral capabilities of each
individual seem rather simple, compared to the global behavior of the group.
Such societies are highly decentralized and often the rules which determine
the interactions among conspecifics are rather simple and locally limited.
Especially for insect societies it is well known that communication heavily
relies on implicit information sharing: Individuals communicate by modifying
their local environment, a mechanisms known as stigmergy. Decentralization and
locally limited interactions make such societies highly robust and flexible to
disturbances, like predation, individual failures, and environmental changes.
In this project we investigated concrete examples which realized a decentralized control approach in the context of evolutionary robotics experiments. We studied why and how distributed control facilitates robustness and resilience to individual failures compared to centralized organizations. We demonstrated how evolution develops control systems which heavily integrate feedback loops with the environment and how this results in a surprising simplicity at the individual neural control level. Eventually, we showed how independent autonomous agents interact with each other to cooperatively accomplish a global function even though they lack the ability to directly communicate with each other. |
|
|
|
||
| movies |
|
Decentralized control architectures work robustly on the real robot... |
|
... as well as in simulation where they were evolved. | |
|
Don't forget: Robots never sleep!!! | |
|
|
||
| related publications |
|
|