(This is a series of short ‘previews’ of papers to be presented at the upcoming Experimental AI for Games workshop at AIIDE 2014. Tune in live on Twitch on October 8th to catch the presentations of these papers, or find the PDFs online at http://www.exag.org)
AI is deeply connected to gameplay, perhaps more than graphics, audio, or other in-game assets. Yet we’ve seen few games that put interaction with AI systems at the core of the game. Existing game AI developed in support of already popular genres like first-person shooters or real-time strategy games. This lead to refined systems for reactive gameplay situations. Classical AI, however, is best at using expressive formalisms for tasks like complex problem solving and question answering. In his paper “Game Design for Classical AI” Ian Horswill designs new game mechanics around high-end classical AI. What problems does an AI-heavy game need to address? What game design supports this kind of AI? Read on for a preview.
Horswill identifies three challenges facing any AI-based game:
- Fragility – AI systems are often sensitive to conditions and can do stupid things when those conditions are not met.
- Transparency – Players need to understand why an AI system does what it does to predict what the AI will do.
- Controlling player expectations – Gameplay must teach players the abilities and limitations of the AI system to play the game.
The paper discusses a variety of mechanics to address these challenges in the following sections.
Classical AI excels at two types of mechanics: problem solving and question answering. Players and AI agents can take many different roles in solving a problem. The paper discusses several AI system roles, the systems used, and examples of games with AI in these roles. Dialog is the core mechanic of interactive fiction and visual novels. The paper discusses question answering mechanics in these games and ways to use AI to extend these systems.
What kind of game can address the three AI system challenges while creating novel mechanics? The paper presents MKULTRA, a game with the premise that the US CIA’s mind control experiments in the 1950s through 1970s succeeded. MKULTRA’s core mechanics are dialog and the use of mind control. Players use freeform text input to influence a non-player character’s (NPCs) stream of consciousness. Commands can gather information from the NPC, give advice to the NPC, or drive NPC conversation. Player can use mind control to give NPCs beliefs to alter their behavior and use telepathic items for effects at a distance or delay. The paper describes ways to address user interface issues to guide text input and display NPC state. After covering the mechanics a technical section breaks out the AI for NPCs to problem solve in the world and for text parsing.