Abstract
Scientific workflows bring together the logistics of complex e-science applications and software services in a user friendly environment. This has become more pronounced with the porting of such applications to the Grid. We report on the design and implementation of scientific workflows to investigate game learning using two of the most widely used scientific workflows systems, Taverna and Kepler, as well as a workflow portal, WS-PGrade. Besides the engineering insight, we offer experimental evidence on how one can set up complex computational processes for investigating interactions between autonomous players, with the final aim of designing optimized interaction sequences for machine learning of playing tactics.
TopIntroduction
Grid computing is about sharing the computing and storage elements of available heterogeneous computer systems to create a pool of resources that would match those of a powerful super computer (Foster, 2002). It is about providing the means to execute computational experiments or applications that, due to their complexity, often require a large amount of computer (storage and computing) resources.
It is the complexity of modern applications and experiments that lead to the development of scientific workflows. These are mostly used to describe, manage and share complex scientific analyses. Different types of tasks described by a workflow can be executed by employing local and/or remote Web services, scripts and sub-workflows (workflows used as fragments of larger ones). Most workflow systems offer a reasonably intuitive interface for setting-up, organizing, submitting and analyzing complex experiments; this interface usually combines a graphical metaphor to model the experiments and a formal language to describe that model. Workflow systems allow users to integrate external applications or services into workflows, deploy the workflows to a grid infrastructure, check for possible errors and finally, analyze the output of these calculations (Taylor, 2007), with relatively modest demands on a user’s IT skills.
Bioinformatics has been one of the scientific fields that generated the demand for grid computing and scientific workflows. As an example, consider in silico studies of biomimetic enzymes, which employ protein structure databases and replica-exchange molecular dynamics algorithms; there has long been a huge demand for powerful computer resources to execute experiments and for sophisticated analysis tools to cope with the large amount of data generated (Dongarra & Rarhavan, 2000; Caragiannis 2013).
We have identified the domain of machine learning in games as a field where grid computing and workflows hold potential; therein, the distribution of computations that are required for the evolution of learned behaviors and strategies must be coupled with a well designed sequence of learning experiments and the accompanying data analysis (Kalles & Fykouras, 2010; Kalles & Kanellopoulos, 2001). We have thus oriented our contribution towards the design and implementation of scientific workflows to investigate game learning using a variety of state-of-the-art tools for workflows, both at the desktop ant at the grid. Based on this experience, we argue that these tools, especially the grid-based one, are of paramount importance to researchers who need to test complex hypotheses that demand excessive amounts of computations. We offer experimental evidence for this argument by showing how the setting we have used can be further exploited for designing optimized interaction sequences for machine learning of playing tactics in gaming.
The rest of this contribution is structured as follows. We first briefly describe scientific workflow software and then we review our initial implementation of experimental sessions in this software. Following that, we review recent developments on workflow systems for grid applications and we set out how to use these resources to analyze the behavior of game playing learning agents. The last section concludes our contribution and identifies promising and ongoing research directions.
Key Terms in this Chapter
Board Game: a game where action takes place on a board (mesh, grid), usually with pawns, where rules typically cover aspects of game playing such as how pawns may move, how players take turns in playing, when the game is deemed to have finished.
Neural Network: an architecture of inter-connected, relatively simple, computational nodes (neurons), where the strength (weight) of the connections allows various input vectors to be mapped to outputs in an relatively arbitrary fashion, thus facilitating classification tasks.
Scientific Workflow: a process model useful for streamlining computations for scientific purposes, akin to a data-flow diagram in structured design.
Reinforcement Learning: a machine learning technique whereby actions are associated with credits or penalties, sometimes with delay, and whereby, after a series of learning episodes, the learning agent has developed a model of which action to choose in a particular environment, based on the expectation of accumulated rewards.
Grid Computing: the infrastructure and services associated with the operation of heterogeneous, distributed computing systems that are primarily targeted at CPU intensive applications which admit parallelism.