Monthly Archives: October 2011

If Words Were Swords

Computationally. ELIZA and its contemporaries are programs that take textual input (natural language) and produce relative, natural language output that only makes sense if it either directly references the user’s previous statement as another statement or poses a question that is at least indirectly related. Removing the textual and language aspect of this model, it is a program that produces output that is somewhat relative to and in the same format as the input that is most optimal, within the bounds of the user’s input and the program’s architecture.

If words were replaced with swords, however, the situation becomes the following: a program that takes combat input (blocking, lunging, etc.) and produces counter or defensive output that is most optimal for the survival of the character (the computational avatar, or more commonly AI). For this instance of a computational artifact, Elder Scrolls IV: Oblivion can compare to ELIZA {her, it}self with its open-world inter-character interactions, which includes combat and ELIZA-like speech with the player character and non-player characters (NPCs).

The greatest computational strength and analog to ELIZA is combat, whereby the characters (player or not) have different attack, defense, and utility abilities. ELIZA keeps a more defensive stance than most Oblivion characters during action, preferring to clearly step around the user’s input and, like a psychotherapist, attempt to lead the user to its own conclusions. The trouble with this, however, is that ELIZA does not quite know when to stop talking and will end up in a loop of canned responses like “why do you say [user input]?” or “could you elaborate on that?” until the user types something that ELIZA is more familiar with handling. These canned, rerouting responses are more of a direct confrontation to the user to try to keep the center of conversation in balance. Compared to Oblivion combat, this is essentially ELIZA’s only attack option aside from dodging deadly blows of questions like “What do you do?” that enter into the aforementioned loops.

Oblivion’s characters are far more impressive with what they do in particular situations, unlike ELIZA, such as waiting for the character to drop their defense (which could have been set up by the NPC with an armor-penetrating attack in the first place) before doing a risky high-damage attack like lunging or drinking a health potion. The fact that the characters keep track of their own status as well as the player is a convincing facet of the AI. If the game were a multiplayer game where other players could control AI, it would indeed be difficult to distinguish between a good instance of AI and a real player.

Convincing the user that his opponent (or chat partner) is a real human is about distinguishing whether that other character is using human-like behavior that the user himself is using. For example, it would be natural for a player, low on health, to run away from an overpowered boss and gulp a health potion or similar utilitarian action instead of running into the fight haphazardly. This is apparent in Oblivion when observing the habits of different computercontrolled opponents (like wolves compared to bandits). Upon sight of the player, the wolf will almost invariably charge the player and never back off; however, a bandit will more likely keep his shield up, cast a few spells, and attack more like he has something to lose.

However, ELIZA, like the wolf, seems to have nothing to lose and will not let up its constant dodging of direct questions (especially ones not explicitly handled in its code) and occasional rerouting counter-questions to keep the focus off of it and onto the user. Whereas Oblivion characters can die or flee, ELIZA has no termination point. Credibility to an artifact’s intelligence may last, but after enough input, patterns or repetitive output occurs and can hurt that credibility.

Most NPCs in Oblivion have player-like interaction abilities as well as tasks. These tasks may be habits like walking to the Mages Guild or jobs like making potions; while mundane, they are the only scripting the characters get for behavior, and interaction with other NPCs and the player determines how and if these goals are accomplished. For example, an Oblivion NPC named City-Swimmer in the in-game city of Bravil is a thief and has thief-like tasks when she does not have food. She will wait until there are fewer other NPCs in the area and then attempt to steal from another NPC. When City-Swimmer or the player commits a crime such as this, NPCs who witnessed the event with report it to the guards and take action themselves; the guards then pursue the perpetrator and swiftly deal justice as they are tasked with themselves. In such combat situations, the characters use input and optimize their output based on their own parameters (peaceful, aggressive, sword, magic). Interestingly, if the player leaves food near her (simply dropped on the ground), she may eat that instead of finding an NPC to pickpocket.

In these cases of the computational artifact as a character’s intelligence, or at least responsiveness to a situation, the options for other characters to act upon are almost always to accomplish its given, often conditional, task. The real action is in the digital interaciton between the NPC and the character in Oblivion; whereas, ELIZA’s constant questions and dodging pushes most of the action onto the user to handle. Oblivion’s AI affords a greater level of immersion and credibility when there are so many options and outcomes of interaction with a single character or even the whole game world compared to ELIZA, which runs into so many unintuitive and rather unintelligent dead-ends to conversation. Perhaps computational artifacts such as chatterbot scripts or game AI that attempt conversation fail because they try to emulate humans’ most unique action, and if a human is to judge its credibility, then it is only a matter of time until the artifact yields too inhuman a response.