Apr 9 2019
For a number of years, Professors Thomas Müller and Hans Briegel have been performing studies on a machine learning model.
Professor Thomas Müller (Image credit: Ulrike Sommer)
The duo’s model is considerably different from the learning models based on alternative artificial intelligence (AI). The philosopher from the University of Konstanz and the theoretical physicist from the University of Innsbruck combine techniques of quantum optics and philosophical action theory. The “Projective Simulation” learning model of the team has already been effectively used in fundamental studies. The researchers, in collaboration with the Innsbruck physicist Dr Katja Ried, have now modified this AI model for realistic uses in biological systems. In the scientific journal PLoS One, the current issue details how the learning model can be utilized for modeling and replicating the particular swarming behavior of locusts.
Demand for models that are “closer to biology”
In order to perform their interdisciplinary collaborative study, the researchers used the data available on locust behavior from the Cluster of Excellence “Centre for the Advanced Study of Collective Behaviour” in the University of Konstanz, which conducts internationally leading studies on collective behavior, and since the beginning of 2019, it is being funded via the German Excellence Strategy.
More specifically, biologists are insisting that models elucidating the overall behavior should be created to be “closer to biology.” A majority of present-day models were developed by physicists who believe that a physical force influences interacting individuals. Consequently, these models do not essentially perceive individuals within swarms to be agents, but rather as points, like interacting magnetization units on a grid.
The models work well in physics and have a good empirical basis there. However, they do not model the interaction between living individuals.
Thomas Müller, Professor, University of Konstanz.
AI rules allow agents to learn
It was Hans Briegel who originally developed the learning model called “Projective Simulation,” which is predicated on agents that do not respond to events in a pre-programmed manner, but rather, they are capable of learning. “Learning agents” like these are coded as individuals with varying behavioral dispositions who perceive and react to sensory input to interact with their environment. For this intent, the learning agents follow AI rules that enable them to apply their earlier individual experiences to modify their actions.
While this learning process involves random processes based on quantum physics during which all possible courses of action are taken into account, the action theoretical principle of reinforcement learning, which is predicated on rewarding specific outcomes, comes into play.
We give a reward if the agent moves with the others in a well-ordered manner. In time, an agent realizes: when perceiving certain things, it is better to react in a way that will lead to a reward. We do not preset the right course of action in a particular situation, but we do ensure that it is achieved through the interaction between the agents.
Thomas Müller, Professor, University of Konstanz.
Learning model can reproduce collective behavior
Hans Briegel, who presently holds a three-year visiting professorship at the Department of Philosophy at the University of Konstanz, Thomas Müller, and Katja Ried have applied this learning model to a particular and well-explored swarming behavior of a locust insect. In a constrained space, the movement behavior of the insect corresponds to the swarm’s size. For example, the locusts move in a disordered manner if there are just a few individuals; however, they move together as a unit in larger numbers. Similarly, they move as a unit and also in the same direction in extremely large numbers. The investigators were first interested in merely testing their learning model, and as a result, they used a qualitative description of the behavior of the locusts instead of raw data. In this fashion, they were actually able to simulate the locusts’ behavior in a qualitative manner.
Looking ahead, Thomas Müller hopes that upcoming studies in this field will gain from huge data sets available on animals, like schools of fish whose behavioral patterns are highly dynamic.
Modelling fish would probably be a good but also very complicated next step in making our learning model even more realistic.
Thomas Müller, Professor, University of Konstanz.