Collaborating and Learning Predators on a Pursuit Scenario

Abstract

Abstract.A generic predator/prey pursuit scenario is used to validate a common learning approach using Wilson's eXtended Learning Classifier System (XCS). The predators, having only local information, should independently learn and act while at the same time they are urged to collaborate and to capture the prey. Since learning from scratch is often a time consuming process, the common learning approach, as investigated here, is compared to an individual learning approach of selfish learning agents. A special focus is set on the performance of how quickly the team goal is achieved in both learning scenarios. This paper provides new insights of how agents with local information could learn collaboratively in a dynamically changing multi-agent environment. Furthermore, the concept of a common rule base based on Wilson's XCS is investigated. The results based on the common rule base approach show a significant speed up in the learning performance but may be significantly inferior on the long run, in particular in situations with a moving prey.Keywords: Multi-agent learning, predator/prey pursuit scenario, emergent behavior, collaboration, and XCS. MotivationDue to the increasing scale and complexity of strongly interconnected application systems there is a need for intelligent distributed information processing and control. The design of multi-agent systems (MASs) has addressed this need, using concepts from machine learning and distributed artificial intelligence [1]. MASs have been utilized successfully in a range of application scenarios: Guiding automated machines in collaborative industry scenarios [2], trading energy on market platforms [3], seeking smallest distance routes for delivery services [4], or managing air conditioners in buildings [5], are some examples of problems which are solved (completely or partially) using MAS approaches.A MAS consists of a collection of agents acting autonomously within their common environment in order to meet their objectives. They take sensory inputs from the environment, match them on actions, and then perform some actions, M. Hinchey et al. (Eds.): DIPES

Adapting the Predator-Prey Game Theoretic Environment to Army Tactical Edge Scenarios with Computational Multiagent Systems

“…Many multiagent research efforts utilize the original discretized predator-prey pursuit environment shown in Figure 1, implementing a toroidal grid world and requiring the single prey to be blocked on all sides by four predators [7,21,22]. These studies tend to focus less on the structure of the environment and more on how specific sets of predator strategies impact cooperation and teamwork in homogenous and heterogeneous groups of predators.…”

“…Some research allows the agents to be omniscient, where all entities' positions are known at every time point [17,35], whereas other studies permit agents to see in a straight line until an occlusion is encountered [33,[38][39][40]. Some allow predators and prey to see only within a given range around themselves [21,30,41] or apply random limitations to predator vision [42]. While some research explores predators with limited sensing abilities, these studies allow information to be shared among predators [21,36], either through direct communication (e.g., police radio, [36]) or indirectly (e.g., by leaving cues such as pheromone trails in the environment) [26].…”

“…Some allow predators and prey to see only within a given range around themselves [21,30,41] or apply random limitations to predator vision [42]. While some research explores predators with limited sensing abilities, these studies allow information to be shared among predators [21,36], either through direct communication (e.g., police radio, [36]) or indirectly (e.g., by leaving cues such as pheromone trails in the environment) [26]. Other work has explored an imbalance between sensing abilities of the predators and prey.…”

“…Other work has explored an imbalance between sensing abilities of the predators and prey. For example, a predator may see a prey from a greater distance than the prey can detect, reflecting a more realistic scenario where the predator is on the hunt for a non-vigilant prey [21]. Other work has further modified agent vision or sensing abilities by testing something akin to sound, where a predator agent can hide around corners [43].…”

“…The original predator-prey pursuit environment ( Figure 1) required four predator agents to surround the single prey agent from four directions in a discretized grid world [1,9,21,23]. The predator agents were guided by an algorithm and their movements were limited to one grid square per time step to an adjacent available square (not occupied by another agent or a boundary) in only the vertical or horizontal directions (no diagonal movements).…”