[Josh Bongard] [Alastair Channon] [Hod Lipson] [Colm Massey] [Tom Ray] [Adam Rotaru-Varga] [Tim Taylor]
[Workshop Home Page] [Workshop Schedule]
Title: Applying Developmental Genetics for Evolving Complete Agents
Recent advances in evolutionary developmental biology have uncovered a host of mechanisms that may prove of use for evolving both the morphology and neural architecture of virtual agents. By incorporating ideas from homeotic genetics, modular agent body plans---complete with distributed neural architecture---can be evolved. It is argued that such an approach to evolving morphology is computationally tractable, and useful for porting agent designs to real-world robots, as well as improving on the benchmark work in this area, conducted by Karl Sims. The application of this work to modular robotics will also be discussed.
Title: The Importance of Brain-Body Coevolution in the Natural Selection of AI-Life
This presentation explores some of the limiting factors identified in my experiments into the natural selection of AI-Life. From two conclusions, each concerned with separate shortcomings in the results, two main issues facing such research can be separated out. In both cases brain-body coevolution presents a promising resolution.
Conclusion 1 (from my ALife VI paper): “This work has made it clear that the specification of 'actions’, even at a low level, results in the organisms being constrained around these actions and limits evolution. Alternatives in which the embodiment of organisms is linked to their actions need to be investigated.”
By using natural selection, as opposed to artificial selection, one can aim beyond the evolution of simple prescribed behaviors, toward the open-ended evolution of emergent behaviors. However, if only the “brains” of agents can evolve, then the range of resulting behaviors will be limited by the body design. This is especially true at the current (early) level of research into AI-Life, where we should expect most evolved behaviors to be motor-based: following, fleeing, circling, flocking, running, jumping, etc.. By allowing bodies to coevolve with their brains, the range of possible motor-behaviors could cease to be a limiting factor. So resolution 1 is to use brain-body coevolution to help make the natural selection of AI-Life open-ended.
Conclusion 2 (from my SAB ’98 paper): “Whether or not emergence is continuing in Geb [the artificial world in question] is hard to tell, for it soon becomes difficult to identify behaviors. This was a less significant problem in the evolution of program code but evolved neural networks are hard to understand and so offer little help. Constructing systems such that behaviors will be more transparent is likely to be the most productive way forward. … alternatives in which the evolvable embodiment of an organism gives rise to its actions will need to be considered.”
The evolution of AI-Life must involve emergent behaviors; It is not be possible to simply specify increasingly complex behaviors. So the problem of understanding, or even identifying, novel behaviors arises. With relatively simple agents, we can analyze their controllers (programs, neural networks, or other) directly to determine the resulting behaviors. But as the complexity of evolved controllers has increased, this has become less and less possible. So, until the agents themselves can help us understand what they are up to (a very distant prospect), we can only observe the resulting behaviors and attempt to identify innovation. An embedded presence, and more so a body, can provide us with the ability to observe the resulting movements of our agents. The logical aim is therefore to allow a range of body possibilities that will make it easier for us to analyze behavioral observations. So resolution 2 is to use brain-body coevolution to aid in the observation (identification and analysis) of emergent behaviors.
The question of how to achieve these two resolutions remains. What degree of freedom in body evolution is required for the open-ended evolution of AI-Life? How close should the physics of an artificial world be to our own, in order for us to best observe emergent behaviors? Thanks to Karl Sims’ artificial selection of “blockies”, we do at least have an idea of where to start.
Title: Issues in Brain-Body Coevolution
I want to raise some issues about body brain coevolution: when can it be considered CO-evolution? Is the destinction between body and brain arbitrary? Is it necessary/sufficient for reaching complexity? etc.
Title: Actuating Virtual Robots
When physically simulating robots, there are a variety of methods at one's disposal to drive or actuate it's degrees of freedom. What are the costs and benefits of each of these techniques, and which are more appropriate when the morphology of the robot is plastic, either within the lifetime of the robot or over generations (in an evolutionary system).
Colm Massey is head of the Intelligent Control Group at MathEngine PLC. Mathengine's current core business is fast rigid body simulations for low-end platforms. The role of the ICG is to develop techniques for controlling these simulations. We have build control systems for a wide variety of simulations, from 'Blocky' like evolved swimmers to self balancing bipeds.
Title: Aesthetically Evolved Virtual Pets
Aesthetic, emotional, and empathetic selection are applied to a derivative of Karl Sims' Evolved Virtual Creatures. The resulting Creatures can be beautiful or strange, and provoke strong reactions in human observers. It may be possible to evolve virtual pets to which humans can form strong emotional bonds.
Title: Theoretical and Encoding Issues
In the first part of my talk I discuss some theoretical issues of brain and body coevolution. Trying to answer the question "Why do we want to do it?" I identify common evolution of control and morphology as a necessary means for evolving creatures with complex (intelligent) behaviours. This is in line with the embodied intelligence view, which recognises the computational role of the body (in addition to the brain's). I identify and describe the so called 'matching problem' as the principal obstacle preventing effective evolution. This problem arises because individual brain components must be connected to individual body components. Different encodings of these two can be disruptive to the matching (and evolution). The merits of variable size, common, and developmental encodings are summarised.
The second part of the contribution provides an overview of the encoding scheme used in the Framsticks system. It seamlessly encodes the brain and body of stick creatures with any tree topology, and includes features like attribute propagation, relative references, and suitable crossover and repair genetic operators. An alternative, a developmental encoding is also mentioned.
Title: New Features in the Latest Physics Engines
I will give a brief overview of the latest offerings from MathEngine and Havok.
Document last updated: Tim Taylor, Thursday, 10 August 2000 15:27:16