IJCAI 2005 Workshop on Reasoning, Representation, and Learning in Computer Games Interfacing the D’Artagnan Cognitive Architecture to the Urban Terror First-Person Shooter Game Bharat Kondeti, Maheswar Nallacharu, Michael Youngblood and Lawrence Holder University of Texas at Arlington Department of Computer Science and Engineering Box 19015, Arlington, TX 76019 {kondetibharat,mailbackmahesh}@yahoo.com, {youngbld,holder}@cse.uta.edu Abstract logical graph from the UrT world maps and uses the graph to find paths between the agent’s starting and goal location. The D’Artagnan Cognitive Architecture (DCA) is a While there is a large body of work in representations and multi-agent framework that supports the study of algorithms for path planning in many domains (e.g., see attention as a means to realize intelligent behavior O’Neill’s [2004] work on a mesh representation for game by weighting the influence of different agents as worlds to support path planning), our work is unique in its they collectively determine the next action. We ability to automatically generate a topological graph from have interfaced the DCA to the Urban-Terror (UrT) the UrT’s map representation, which is the means by which first-person shooter game and defined several different UrT world scenarios are distributed to gamers. worlds of increasing complexity in order to test the To test the DCA approach, we define multiple tasks in DCA’s ability to perform well in these worlds and five different UrT maps and evaluate the performance of a demonstrate the usefulness of shifting attention reflex-agent-based DCA while playing UrT. Our goal is to among different agents. We have implemented evaluate the hypothesis that the DCA consisting of multiple several reflex agents and a path planner to help action-generating agents, controlled by a global attention DCA play the UrT game. Experimental results in- agent, can accomplish tasks too difficult for a single-agent- dicate that a DCA-based player using a combina- based DCA. This hypothesis is similar to that confirmed in tion of action-determining agents can be successful Reynolds’ [1999] work, where he exhibited human- when no single agent can complete the task. consistent steering behavior using a linear combination of numeric values from lower-level behaviors (e.g., flee dan- 1 Introduction ger, avoid obstacles). However, the DCA must choose Interactive computer games have been considered human- among a discrete set of actions for which we propose an level AI’s “killer app” [Laird and van Lent, 2001] in that approach based on an adaptive focus of attention. current games have a sufficient level of realism to require human-level intelligence to play well. Laird and van Lent’s 2 D’Artagnan Cognitive Architecture (DCA) work along these lines with the SOAR cognitive architec- The D’Artagnan Cognitive Architecture (DCA, ture and the Unreal Tournament game explored the current http://ailab.uta.edu/dca) [Youngblood, 2000; Youngblood limits of AI to play these games [Laird, 2002]. Motivated and Holder, 2003] is based on the work of existing cognitive from this challenge, but with an alternative view of the de- architectures, robotics research, and human-consistent cog- sign of cognitive architectures, we have begun development nitive models, centered on a specific task. DCA consists of on the D’Artagnan Cognitive Architecture (DCA) and an twelve components, or models of cognition (see Figure 1). interface between it and the Urban Terror (UrT) first-person The twelve DCA models consist of the action model, action shooter game. The DCA is a novel approach to cognitive evaluation model, attention model, decision model, effectual architectures based on a Minskian society-of-agents ap- model, emotion model, goal model, learning model, learning proach [Minsky, 1988] of psychologically-inspired, possibly evaluation model, memory model, perceptual model, and competing, agents with a global focus-of-attention influence reflex model. Models are implemented using one or more over the agents to achieve robust, human-consistent intelli- agents in a multi-agent framework, e.g., several different gent behavior. In previous work we have presented metrics types of learning techniques may be underlying the learning for human-consistency and comparison of human and DCA model, and all compete for the bandwidth to influence behaviors [Youngblood and Holder, 2003]. DCA’s behavior. When connected, these components form In this paper we describe the DCA, the UrT game, and an architectural system for learning and adapting to an envi- the interface between the two. Fundamental to a player’s ronment. Figure 1 depicts one possible set of connections performance in such environments is the ability to reason between models, but in reality the models are completely spatially. Therefore, we have also implemented a path plan- connected. The communication bandwidth across a connec- ner based on the work of [Hill, 2002] that generates a topo- Learning Memory Perceptual Evaluation Model Input Model (Sensors) Learning Attention Reflex Model Model Model Goal Decision Action Model Model Model Effectors Emotion Action Model Evaluation Model Figure 2. Urban area map (left) used by the DCA- Figure 1. The D'Artagnan Cognitive Architecture (DCA). UrT project and a screen shot of the game (right). implemented an interface to the UrT game that allows the tion varies dynamically, as controlled by the attention DCA (or any other system) to extract perceptual information model. from the UrT game and perform actions in UrT. UrT sup- When percepts are received and stored into memory, the ports several challenging world maps (e.g., Figure 2 depicts appropriate models are triggered and start processing. Some an UrT urban map and a game screenshot) and game scenar- models will select goals and propose actions based on learn- ios (e.g., capture the flag, bomb-defuse, and free for all). We ing from past actions, while other models will use delibera- have also defined our own simplified games within these tive planning to determine the next action leading toward a worlds that still portray realistic urban warfare scenarios. goal. The proposed actions are available at any time for the We have developed mechanisms for logging game informa- action model, which selects the action to take. The selected tion in XML in order to extract a player’s behavior. Eventu- action is executed by the effectors. ally, we plan to have human players play our worlds in or- Selection among the set of possible actions is affected by der to capture their play, which will serve as part of an a number of factors. The learning model not only generates evaluation metric for an agent’s consistency with human knowledge and possible actions, but also evaluates past de- behavior. cisions to learn the action's effectiveness for a particular goal. The emotion model provides a suppressor or enabler signal of an action based on environmental situations. 4 DCA-UrT Interface Strong emotions can completely inhibit certain actions and The DCA and UrT are interfaced via shared memory to ex- enable others (e.g., fear will inhibit charging an opponent change percepts and actions. The shared memory is used to and enable retreat). The final dispatcher of influence is the read and write percepts and actions with lower communica- attention model, which controls the weights of all edges and tion latency and lower computational burden on the game thus controls communication between models and the confi- engine. A visual interface called UrTInterface (see Figure 3) dence level of possible decisions generated by these models. has been developed for UrT to display all the percept infor- The attention model also controls the timing of decisions, mation that can be obtained from the game and also acts as a following the anytime paradigm, to produce the best possi- virtual keyboard to play the game. This section describes the ble decision at given time intervals. Based on the human main aspects of interfacing DCA to UrT. For more interface ability to focus the mind on different thought mechanisms details, see [Kondeti, 2005]. for different situations, the attention model can stop and restart other models to enforce desired behavior. 4.1 Modifications to UrbanTerror Since UrbanTerror (UrT) is a realistic shooter game, with 3 Urban Terror (UrT) sophisticated worlds and adversaries, and DCA is still in its We have begun development on interfacing DCA to a visu- infancy, a number of modifications had to be done to UrT ally and tactically realistic urban warfare simulator called before DCA can play it. The main concern of DCA is to Urban Terror (UrT, http://www.urbanterror.net), which is navigate towards the goal while avoiding obstacles. So the built on top of the cross-platform (Win32, Linux, Mac OS opponents in the game had to be removed since their goal is X) Quake III Arena game engine. UrT is a first-person to kill the player and end the game. This is done by deliber- shooter (FPS) game developed by Silicon Ice Development. ately not allowing the game to start any opponents in the At present UrT is offered as an entertainment-based game, game, giving DCA more time to successfully navigate to the and to our knowledge, has not been deployed for any other goal without getting killed by opponents. commercial or military use. As part of this project we have The goal of the DCA is to get the opponent’s flag. The rules for “Capture the Flag” mode in UrT require the player Figure 3. DCA-UrT effector interface (left) and percept interface (right). to get to the opponent’s flag and return back to his base flex agents and selects an action based on a policy. Since without being killed or losing the flag. Since there are no there are no agents to evaluate the action taken, the policy is opponents and the task for DCA to navigate toward the goal hard-coded into the action agent.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages6 Page
-
File Size-