Experiments with a Social Learning Model

Abstract

We present a system that models perception-action coupling through imitation and attention. Our interest is in imitation and in social learning more generally. Through social learning the experience of an agent is governed by the actions of an expert, and the structures that develop within the agent's "brain" are influenced by its social situatedness. We are inspired from biological findings in primates of the existence of mirror neurons, which are believed to be involved in imitation. The visual and motor properties of these neurons suggest a tight perception-action coupling, where affordances could be expressed. Our system is designed to model the functional properties of the mirror neurons and therefore express the functionality of objects. The system builds up perceptual and motoric structures from experience using temporal attention and forms perceptual-motor connections. The experience arises through imitation, where an agent can perceive objects and the interactions upon them. We have successfully applied our system on three different platforms, two in simulation and the third on a real robot learning from a human. The system is able to segment the perceptual-motor experience into distinct structures that can be used to recognize and reproduce the task in each experiment. Some unexpected results showed us that the motoric complexity in these experiments was not high enough to expose the full potential of our system, and we suggest future work that will address these results.

How Game Complexity Affects the Playing Behavior of Synthetic Agents

“…Connect-4 is a turnbased game and each agent has exactly one move per turn. 1,2,3,4,5,6,7,8,9,10 It has a medium state space complexity of 4.5 × 10 12 board positions [16]. Fig.…”

Synthetic learning agents in game-playing social environments

“…For a game agent, the social environment is represented by a game with all its agents, components and entities, such as rules, pay-offs and penalties, amongst others (Ferber, 1999; Kiourt & Kalles, 2012). Learning in a game is said to occur when an agent changes a strategy or a tactic in response to new information, thus mimicking human playing behaviour (Al-Khateeb & Kendall, 2011; Caballero et al, 2011; Elidrisi, Johnson, Gini, & Crandall, 2014; Gilbert & Troitzsch, 2005; Kiourt & Kalles, 2012; Lopes et al, 2009; Marivate & Marwala, 2008; Marom et al, 2001). It is widely held that games are capable of simulating human game-playing behaviours, such as reactions, defeat, learning speeds, strategy options, etc.…”

“…Simulation introduces the possibility of a new way of thinking about such processes, based on ideas about the emergence of complex behaviour from relatively simple activities (Ferber, 1999; Gilbert & Troitzsch, 2005; Poole & Mackworth, 2010). Simply put, social learning is concerned with how humans learn from observations of others’ actions (Ferber, 1999; Marom, Maistros, & Hayes, 2001; Lopes, Melo, Kenward, & Santos-Victor, 2009; Poole & Mackworth, 2010). When many agents interact with each one usually acting selfishly, produce multi-agent systems and social organizations, which attempt to simulate traits of human behaviour (Gilbert & Troitzsch, 2005; Kiourt & Kalles, 2013).…”

Computational techniques for modeling non-player characters in games

“…During the past years, learning by imitation has been a popular method of acquiring complex behaviors [152] [12]. This kind of behavior learning can be applied in very broad areas like recognizing human action sequences [88] [75], creating learning robots [173] [95] and creating humanoid virtual characters [124] [91]. Imitative learning is also useful in solving problems, such as dynamically generating the gaming intelligence according to user's play patterns [168].…”