Evolving Virtual Creatures

Contribution:

Virtual characters are represented as directed graphs having nodes and connection. A genetic language based evolution of motor as well as neural control is suggested. No user information regarding shape/size, joint constraints etc are assumed when developing the creatures.

Computation of results:

Since the entire task is computationally expensive, parallel implementation of the task is done. Evolutions of about 50 to 100 generations i.e. by mating directed graphs either by crossover operation (30%) or by grafting(30%) and sometimes asexually (only using mutations) (40%), are performed corresponding to four types of behavior, i.e. swimming, jumping, walking and following. Another suggested method was to interactively evolve a morphology so as the creatures have aesthetic touch to them instead of just goal based evolution. Hybrid world creatures were also developed.

Reproducibility:

My view is that the paper is reproducible. Although models are not detailed mathematically, they make up for it by giving exhaustive description of the same. The models used for sensors, nerons and Effectors are detailed properly in section 3. Details about the use of physical simulation techniques e.g Feather's method to calculate accelerations, Runge-Kutta fehlberg method for integration and other details for virtual world simulation (e.g water, collison etc) are provided in section 4. The genetic optimization based behavior selection method is also well explained in section 5.

Improvements in Research/Writing:

Overall the paper is well written but I would have preferred some more light to be thrown on the technical aspects of the overall model and specially the modeling of characters as directed graphs and the genetic evolution process.

-- Main.sumanm - 17 Nov 2011

Above is the review of : "Evolving Virtual Creatures"

-- Main.sumanm - 17 Nov 2011

What is the contribution of the paper?

In previous work, control system must be designed for each type of fixed structure. The method in this paper can generate physical structure and its corresponding control system automatically in a similar fashion to creature evolution. So the generated physical structure and the control system is also called a creature. Users don't have to specify any parameters of physical structure or control system. The generated physical structure and control system work in 3D physics.

How are the results evaluated?

The results are evaluated by looking at the generated creatures in various environments. An important concern in the evaluation seem to be how well creatures can adapt to the environment by performing expected tasks, i.e. does it swim or walk fast, or does it jump high?

It is also evaluated by the number of generated creatures and the time it takes to do so.

Is the paper reproducible?

Somewhat reproducible, but it's not clear to what genotypes look like and how they can be converted to creature morphology and creature control.

how could the paper research or paper writing be improved?

Overall, this paper is well written, but in the introduction, there is not an overview of sections two through six and how those sections are related to one another. It would be nice if they do so.

Main.shuoshen - 17 Nov 2011

• contribution: The research creates a novel system for creating virtual creatures behaving in 3D physical world. The creature’s morphologies as well as their control systems are evolved. “A genetic language is presented that uses nodes and connections as its primitive elements to represent directed graphs, which are used to describe both the morphology and the neural circuitry of these creature.”[Karl Sims]

• evaluation: The experiment runs in parallel with one master node performing the genetic algorithm that chosen randomly from asexual, crossovers and grafting, and then sending out genotypes to other slave nodes to do fitness test and then gathering back the fitness values. The fitness tests include swimming, walking, jumping and following. After 50 to 100 generations, creatures that pass the fitness threshold survive and get selected.

• reproducibility: In my opinion the paper should be reproducible however it will involve large amount of work since the models used in the research are presented without detailed mathematics explanation. And I still don’t quite get how the phenotype “brain” is generated. But I think the physical simulation part shouldn’t be that hard since we use simplified models.

• comments: Overall the paper is well organized and I found the structure is not hard to follow. However it takes time to learn some background knowledge.

Baoxuan Xu

Evolving Virtual Creatures - David Matheson

(a) Contribution: The key contribution is a framework for applying genetic algorithms to optimize creature morphologies and controller systems. They represent both with directed graphs where the nodes specify rigid bodies and partial "neural nets" (they're not really neural nets).

(b) Evaluation: The results were evaluated qualitatively. For each task they describe different morphologies and control strategies that were selected. It would have been nice to see the relative performance of different morphology and control schemes.

(c) Reproducibility: The idea of the paper is reproducible and their evaluation functions are well documented. However the parameters for the possible rigid bodies are not specified. Also the parameters for the simulation such as high and low velocities for switching between spring/impulse models are not provided.

(d) Improvements in Research/Writing: This paper is a framework for applying genetic algorithms to morphology and control systems. Different meta control schemes could be evaluated using different models for control. As long as it maintains the requirements of the genetic selection it will still work as part of the framework. I think the paper could have been clearer by specifying the requirements for applying genetic selection to morphology and control independently of their particular specification of morphology and control.

* Contribution: In previous research, the control system can automatically be formed rather than design for specific physical structure. Users don't need to care about the detail shape of the object and it works in 3D space.

* Result Evaluation: This paper use several environments to evaluate the result. The performance time is concerned. And also, this system might work in different jobs, thus they test it by making the model do many kinds of motion and evaluate the physical realistic animation.

* Reproducibility: This paper is OK to reproduce. Maybe it is hard for us to regerate some math methods and those models.

* About the improvement: It seems no fruitful background and relate previous work for use to understand the breakthrough of this paper in a comprehensive view.

-- Main.Jingxian Li - 17 Nov 2011

* What is the contribution of the paper?

This paper proposed a method to use directed graph to represent virtual creatures and use nervous system derived from this to remove user-design phase.

* How are the results evaluated?

First, we create an initial population of genotypes with no hand-designed ones. And then, we use the survival-ratio to determine the percentage of the population that will survive each generation. For each generation, we first mutate the directed graph. There are five detailed steps for mutating. If the directed graph is nested, we first mutate the outer and then the inner. After we mutate, we extend the graph and we need to mate it. There are two ways to evaluating this, the first way is to use crossover operator. And the other way is to graphs two genotypes together. And for the implementation approach, we use parallel computation and prevent idle processors by starting new generation before fitness test finished.

* Is the paper reproducible?

It depends. For the data, it is obviously reproducible. And the algorithm is generic so it is reproducible. But the code is not reproducible because there is even no pseudo-code.

* How could the paper research or paper writing be improved?

To give the real implementation of the algorithm, at least, the key parts. The reproducibility is highly recommended nowadays. And this paper can also divide the introduction part here into introduction and previous work separately.

-- Main.chuanzhu - 18 Nov 2011

The paper contributed by introducing a method for generating and testing simple hierarchical feedback-responsive creatures in a thought-provoking demonstration of structured behavior evolving from utter randomness. This is essentially a step forward towards automated creation of new and interesting solutions fitting a limited set of behavioral constraints set by an animator and letting a computer take care of the "creative" generation. From an AI perspective, this paper demonstrates an alternative towards the hard task of designing intelligent systems by throwing computing power (chaotically, yet wisely) at the problem.

The results were evaluated based on combination of subjective or arbitrary discrimination and optimization of an objective score based on success in a few low-level motion tasks. However, the paper did not describe fully the process for discriminating against evolutionarily slow developers and genuinely poor ones. This appears to leave out the possibility of a beneficially unique behavior (but poor survivability) by chance combining with general survival traits. It is also not well described how two creatures evolved to do two separate tasks can combine to do either or both tasks. These would have been helpful to help shape complex motions useful for an animator.

While the process is described, the paper does not appear to provide enough information for others to easily reproduce the findings. It does not provide sample or reference implementations, and does not provide background for the techniques used for the network implementation on the parallel architecture. It does not even provide the metrics and examples of the objective functions, parameter, or constraints in the cherry-picked creatures shown. While their stance that the idea is not to understand the design process, the images and videos do not tell whether the mechanics behind the behavior are actually practical or meaningful. -- KevinWoo

(a) What is the contribution of the paper?
This paper describes how virtual creatures can be created and evolved. Simulation takes place in three-dimensional space. Shape, size, overall embodiment and the control system of the creatures are represented as a genetic language using directed graphs. A Darwinian "survival of the fittest" is used as a natural selection in process of creatures' evolution, where creatures being tested for best performance of tasks such as swimming, walking, jumping, following the light source. Generations were altered by mutating and mating directed graphs.

(b) How are the results evaluated?
The genetic algorithm was implemented to run in parallel, and an evolution with population size 300 ran for 100 generations. Evolutions were performed separately for each of the tasks. Unlike jumping, swimming, walking and light-following tasks resulted in a large number of creaters with variaty of strategies accomplishing the desired behaviours.

(c) Is the paper reproducible?
I believe so. General ideas of algorithms are presented in detail, therefore the implementation can reproduced by referencing corresponding maths behind it. However, the algorithms concerning evoluation and behaviour performance are not clear enough to be easy reproducable.

(d) How could the paper research or paper writing be improved?
The paper is well written. However, I found 6.2 Mating Directed Graph a bit unclear, because mating methods described briefly and I found them not well reasoned.

-- Main.khamza - 18 Nov 2011