Hypernetworks Analysis of RoboCup Interactions

Rossi, Ruggero (2022). Hypernetworks Analysis of RoboCup Interactions. PhD thesis The Open University.

DOI: https://doi.org/10.21954/ou.ro.000145e2

Abstract

Robotic soccer simulations are controlled environments in which the rich variety of interactions among agents make them good candidates to be studied as complex adaptive systems. The challenge is to create an autonomous team of soccer agents that can adapt and improve its behaviour as it plays other teams. By analogy with chess, the movements of the soccer agents and the ball form ever-changing networks as players in one team form structures that give their team an advantage. For example, the Defender’s Dilemma involves relationships between an attacker with the ball, a team-mate and a defender. The defender must choose between tackling the player with the ball, or taking a position to intercept a pass to the other attacker. Since these structures involve more that two interacting entities it is necessary to go beyond networks to multidimensional hypernetworks. In this context, this thesis investigates (i) is it possible to identify patterns of play, that lead a team to obtain an advantage ?, (ii) is it possible to forecast with a good degree of accuracy if a certain game action or sequence of game actions is going to be successful, before it has been completed ?, and (iii) is it possible to make behavioural patterns emerge in the game without specifying the behavioural rules in detail ? To investigate these research questions we devised two methods to analyse the interactions between robotic players, one based on traditional programming and one based on Deep Learning. The first method identified thousands of Defender’s Dilemma configurations from RoboCup 2D simulator games and found a statistically significant association between winning and the creation of the defender’s dilemma by the attackers of the winning team. The second method showed that a feedforward Artificial Neural Network trained on thousands of games can take as input the current game configuration and forecast to a high degree of accuracy if the current action will end up in a goal or not. Finally, we designed our own fast and simple robotic soccer simulator for investigating Reinforcement Learning. This showed that Reinforcement Learning using Proximal Policy Optimization could train two agents in the task of scoring a goal, using only basic actions without using pre-built hand-programmed skills. These experiments provide evidence that it is possible: to identify advantageous patterns of play; to forecast if an action or sequence of actions will be successful; and to make behavioural patterns emerge in the game without specifying the behavioural rules in detail.

Viewing alternatives

Download history

Metrics

Public Attention

Altmetrics from Altmetric

Number of Citations

Citations from Dimensions

Item Actions

Export

About

  • Item ORO ID
  • 83426
  • Item Type
  • PhD Thesis
  • Keywords
  • robotics; soccer; computer simulations; football; artificial intelligence; neural networks (computer science); machine learning
  • Academic Unit or School
  • Faculty of Science, Technology, Engineering and Mathematics (STEM)
  • Copyright Holders
  • © 2021 Ruggero Rossi
  • Depositing User
  • Ruggero Rossi

Recommendations