Axiomatic Scalable Neurocontroller Analysis via the Shapley Value

Keinan, Alon; Sandbank, Ben; Hilgetag, Claus C.; Meilijson, Isaac; Ruppin, Eytan

doi:10.1162/artl.2006.12.3.333

Cited by 52 publications

(61 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It is worth noting that the so-called Shapley value, which is a concept from game theory, has been applied to computational neuroscience, with different brain areas considered as "players" [71]. The Shapley value is used for calculating the contribution of players in a coalition to the overall performance (cf.…”

Section: Discussionmentioning

confidence: 99%

Use of Game-Theoretical Methods in Biochemistry and Biophysics

et al. 2008

View full text Add to dashboard Cite

Evolutionary game theory can be considered as an extension of the theory of evolutionary optimisation in that two or more organisms (or more generally, units of replication) tend to optimise their properties in an interdependent way. Thus, the outcome of the strategy adopted by one species (e.g., as a result of mutation and selection) depends on the strategy adopted by the other species. In this review, the use of evolutionary game theory for analysing biochemical and biophysical systems is discussed. The presentation is illustrated by a number of instructive examples such as the competition between microorganisms using different metabolic pathways for adenosine triphosphate production, the secretion of extracellular enzymes, the growth of trees and photosynthesis. These examples show that, due to conflicts of interest, the global optimum (in the sense of being the best solution for the whole system) is not always obtained. For example, some yeast species use metabolic pathways that waste nutrients, and in a dense tree canopy, trees grow taller than would be optimal for biomass productivity. From the viewpoint of game theory, the examples considered can be described by the Prisoner's Dilemma, snowdrift game, Tragedy of the Commons and rock-scissors-paper game.

show abstract

Section: Discussionmentioning

confidence: 99%

Use of Game-Theoretical Methods in Biochemistry and Biophysics

et al. 2008

View full text Add to dashboard Cite

show abstract

“…To quantify the importance of each synaptic connection in the original neurocontroller, we use a perturbation paradigm [1], afflicting numerous multiple perturbations upon the synaptic connections, while measuring the corresponding performance score of the agent during each such perturbation. Based on these data, the importance of each synaptic connection to the agent's overall fitness can be rigorously assessed via the Multi-perturbation Shapley value Analysis (MSA) [8,7]. We find that the 14 synaptic connections that survived all the 20 different ENM runs are all among the first 20 most important synaptic connections according to this analysis.…”

Section: The Retained Synaptic Connectionsmentioning

confidence: 99%

“…of the intact neurocontroller [8,7]. ENM prunes 96% of the synaptic connections while maintaining the original fitness almost intact, whereas the other methods result in a significant decrease in the agent's fitness already at much smaller levels of pruning.…”

Section: Minimization Of Successful Neurocontrollersmentioning

confidence: 99%

Neurocontroller Analysis via Evolutionary Network Minimization

Ganon¹,

Keinan

Ruppin

2006

Artificial Life

Self Cite

View full text Add to dashboard Cite

This study presents a new evolutionary network minimization (ENM) algorithm. Neurocontroller minimization is beneficial for finding small parsimonious networks that permit a better understanding of their workings.The ENM algorithm is specifically geared to an evolutionary agents setup, as it does not require any explicit supervised training error, and is very easily incorporated in current evolutionary algorithms. ENM is based on a standard genetic algorithm with an additional step during reproduction in which synaptic connections are irreversibly eliminated. It receives as input a successfully evolved neurocontroller and aims to output a pruned neurocontroller, while maintaining the original fitness level. The small neurocontrollers produced by ENM provide upper bounds on the neurocontroller size needed to solve a given task successfully, and can provide for more efficient hardware implementations.

show abstract

“…Although extremely interesting, providing an answer to this type of questions is not always a simple task. The methodologies we have in order to look for the operational mechanisms of evolved neural networks are limited to networks with a small number of neurons, or to cases in which the neural networks control simple agents that can only move in a one-dimensional world, or by discrete steps (see [4,5,26] for details). Due to the nature of our system, most of these methods cannot be directly employed to investigate which mechanisms control the process by which two homogeneous s-bots differentiate into s-bot gripper and s-bot grippee.…”

Section: An Operational Descriptionmentioning

confidence: 99%

Evolving Self-Assembly in Autonomous Homogeneous Robots: Experiments with Two Physical Robots

et al. 2009

View full text Add to dashboard Cite

This research work illustrates an approach to the design of controllers for self-assembling robots in which the self-assembly is initiated and regulated by perceptual cues that are brought forth by the physical robots through their dynamic interactions. More specifically, we present a homogeneous control system that can achieve assembly between two modules (two fully autonomous robots) of a mobile self-reconfigurable system without a priori introduced behavioral or morphological heterogeneities. The controllers are dynamic neural networks evolved in simulation that directly control all the actuators of the two robots. The neurocontrollers cause the dynamic specialization of the robots by allocating roles between them based solely on their interaction. We show that the best evolved controller proves to be successful when tested on a real hardware platform, the swarm-bot. The performance achieved is similar to the one achieved by existing modular or behavior-based approaches, also due to the effect of an emergent recovery mechanism that was neither explicitly rewarded by the fitness function, nor observed during the evolutionary simulation. Our results suggest that direct access to the orientations or intentions of the other agents is not a necessary condition for robot coordination: Our robots coordinate without direct or explicit communication, contrary to what is assumed by most research work in collective robotics. This work also contributes to strengthening the evidence that evolutionary robotics is a design methodology that can tackle real-world tasks demanding fine sensory-motor coordination.

show abstract

Axiomatic Scalable Neurocontroller Analysis via the Shapley Value

Cited by 52 publications

References 34 publications

Use of Game-Theoretical Methods in Biochemistry and Biophysics

Use of Game-Theoretical Methods in Biochemistry and Biophysics

Neurocontroller Analysis via Evolutionary Network Minimization

Evolving Self-Assembly in Autonomous Homogeneous Robots: Experiments with Two Physical Robots

Contact Info

Product

Resources

About