Periodic Reporting for period 1 - CoEvolFramework (Unified Framework for the Analysis of Co-evolutionary Systems)
Periodo di rendicontazione: 2016-02-01 al 2018-01-31
Today's challenges are marked by more frequent and wide-spread episodes of social, economic, political and environmental crisis. Early studies have formulated these real-world problems in the setting of systems of many interacting adaptive agents involving situations of strategic decision-making. Coevolutionary systems are developed to provide natural realizations and powerful tools to understand conditions that affect the emergent (macro) behavior of these systems as populations of agents whose individual (micro) behaviors changes in response to their interaction outcomes. However, current state of the art research is limited by the complex relationships underlying agent interactions (structures) and their effects on the evolutionary change (dynamics) of the system's collective behavior. Studying these complex coevolutionary systems remains an open challenge as rich structures are not taken into account fully.
This project aims to fill this major research gap with a unified, principled framework to analyze such complex coevolutionary systems. At the core of our approach is the directed graph (digraph) representation of interacting agent behaviors (strategies) where problem structures are fully captured by orientations in the complete graph and associated coevolutionary dynamics by sampling processes on the digraph. Our framework allows us to develop an in-depth qualitative understanding of those problem structures and construct tools for precise quantitative characterizations of any coevolutionary system. We are able to directly relate coevolutionary problem complexity arising from underlying cycle structures of its digraph representation and how it affects the dynamics of coevolutionary search processes.
This project aims to fill this major research gap with a unified, principled framework to analyze such complex coevolutionary systems. At the core of our approach is the directed graph (digraph) representation of interacting agent behaviors (strategies) where problem structures are fully captured by orientations in the complete graph and associated coevolutionary dynamics by sampling processes on the digraph. Our framework allows us to develop an in-depth qualitative understanding of those problem structures and construct tools for precise quantitative characterizations of any coevolutionary system. We are able to directly relate coevolutionary problem complexity arising from underlying cycle structures of its digraph representation and how it affects the dynamics of coevolutionary search processes.
"Preparatory work starts with formulating a unified, principled framework for representing coevolutionary systems. The first task starts with the specification of coevolutionary systems: (1) coevolutionary problem - the set of all alternatives (strategies in a two-player game) and a pairwise preference relation indicating the superior alternative (when one strategy defeats another one), and (2) coevolutionary process - the population of alternatives undergoing a process of selection and variation guided only by their interaction outcomes. Such specified coevolutionary systems can be formulated through: (1) digraph representation of coevolutionary problem that fully captures structures in the pairwise relations between strategies, and (2) Markov chain modelling of coevolutionary processes as random walks on digraphs. Then, macro-scale indices for precise characterizations of coevolutionary systems are developed. This involves qualitative characterizations of coevolutionary problems through cycle structures by examining the presence of strong connectivities in the digraphs. Solvable coevolutionary problems are associated with reducible digraphs that can be vertex-partitioned whereby a subset of strategies dominate the rest. Non-solvable cooevolutionary problems are associated with irreducible (strongly connected) digraphs that cannot be vertex-partitioned.
This qualitative result is used on the second task to develop quantitative characterizations of coevolutionary processes modelled as Markov chains. Coevolution of a solvable problem corresponds to an absorbing Markov chain where one can precisely formulate the expected hitting times to the absorbing class (subset of dominant strategies). Coevolution of a non-solvable problem corresponds to an irreducible Markov chain, in which case the quantity of interest is its limiting invariant distribution. Both theoretical and computational case studies have been carried out to demonstrate how the analysis can be performed.
The third task involves quantifications of coevolved solutions generated by the coevolutionary system. The main idea involves introducing random restart to the associated coevolutionary process. One can then calculate the visitation probabilities of vertices (strategies) by the coevolutionary random walker for both reducible and irreducible digraphs. Theoretical groundings and empirical support for such quantifications have been provided in this task. The visitation probabilities correspond to performance measure of coevolved solutions and they are mathematically shown to exist. Changes to those results due to changes in restart probabilities can be quantified precisely.
The fourth, final task involves comparison of dynamics in coevolutionary processes. This requires evaluating outcomes from coevolutionary random walks over a benchmark of coevolutionary digraphs with different levels of complexity. A benchmark of digraphs of known cycle structures from digraph theory and those obtained from a new generative methodology from statistical physics has been developed. The network-growth-based approach can produce a hierarchy of random coevolutionary digraphs where a control parameter is used to move between two complexity extremes (irreducible digraphs and reducible digraphs with prominent transitive structures). Coevolutionary digraph complexity is quantified precisely with a newly developed, digraph-theoretic index that count the number of strong components. Computational studies have uncovered how these structures affect the coevolutionary search process.
The outcomes of first, second and fourth tasks have been published as ""A New Framework for Analysis of Coevolutionary Systems - Directed Graph Representation and Random Walks,"" Evolutionary Computation, MIT Press (https://doi.org/10.1162/evco_a_00218). Outcomes from the third task have been submitted for a high-impact journal publication."
This qualitative result is used on the second task to develop quantitative characterizations of coevolutionary processes modelled as Markov chains. Coevolution of a solvable problem corresponds to an absorbing Markov chain where one can precisely formulate the expected hitting times to the absorbing class (subset of dominant strategies). Coevolution of a non-solvable problem corresponds to an irreducible Markov chain, in which case the quantity of interest is its limiting invariant distribution. Both theoretical and computational case studies have been carried out to demonstrate how the analysis can be performed.
The third task involves quantifications of coevolved solutions generated by the coevolutionary system. The main idea involves introducing random restart to the associated coevolutionary process. One can then calculate the visitation probabilities of vertices (strategies) by the coevolutionary random walker for both reducible and irreducible digraphs. Theoretical groundings and empirical support for such quantifications have been provided in this task. The visitation probabilities correspond to performance measure of coevolved solutions and they are mathematically shown to exist. Changes to those results due to changes in restart probabilities can be quantified precisely.
The fourth, final task involves comparison of dynamics in coevolutionary processes. This requires evaluating outcomes from coevolutionary random walks over a benchmark of coevolutionary digraphs with different levels of complexity. A benchmark of digraphs of known cycle structures from digraph theory and those obtained from a new generative methodology from statistical physics has been developed. The network-growth-based approach can produce a hierarchy of random coevolutionary digraphs where a control parameter is used to move between two complexity extremes (irreducible digraphs and reducible digraphs with prominent transitive structures). Coevolutionary digraph complexity is quantified precisely with a newly developed, digraph-theoretic index that count the number of strong components. Computational studies have uncovered how these structures affect the coevolutionary search process.
The outcomes of first, second and fourth tasks have been published as ""A New Framework for Analysis of Coevolutionary Systems - Directed Graph Representation and Random Walks,"" Evolutionary Computation, MIT Press (https://doi.org/10.1162/evco_a_00218). Outcomes from the third task have been submitted for a high-impact journal publication."
The unified and principled framework for coevolutionary systems has deepened our understanding and knowledge of real-world strategic decision-making systems that are naturally realized as populations of adaptive agents. Unlike past studies, this framework takes into account the full cycle structures associated with real-world problems and can be used to explain how those structures affect the dynamics of coevolutionary processes in the search for solutions.
Beyond the natural and evolutionary computation community, the project has implications and impact to the wider Artificial Intelligence community and practitioners in other fields who study systems involving situations of strategic decision-making. One important issue dealing with problems in competitive settings is how problem structures involving pairwise interactions can affect the performance of solutions and their discovery by iterative procedures. This framework enables the construction of tools for rigorous quantitative analysis of complex coevolutionary systems.
Given that the two-year fellowship has focused on theoretical development, the work has been well-received by the scientific community. One archival contribution has been made at the top theory journal of the field (Evolutionary Computation). Another has been submitted to a high-impact journal publication of a wider scope. Among others, the Fellow received substantial scientific training that enabled him to combine distinct areas of digraph theory and complex, adaptive, and dynamical systems in a novel way to produce these works.
Beyond the natural and evolutionary computation community, the project has implications and impact to the wider Artificial Intelligence community and practitioners in other fields who study systems involving situations of strategic decision-making. One important issue dealing with problems in competitive settings is how problem structures involving pairwise interactions can affect the performance of solutions and their discovery by iterative procedures. This framework enables the construction of tools for rigorous quantitative analysis of complex coevolutionary systems.
Given that the two-year fellowship has focused on theoretical development, the work has been well-received by the scientific community. One archival contribution has been made at the top theory journal of the field (Evolutionary Computation). Another has been submitted to a high-impact journal publication of a wider scope. Among others, the Fellow received substantial scientific training that enabled him to combine distinct areas of digraph theory and complex, adaptive, and dynamical systems in a novel way to produce these works.