Evaluation

The most significant addition to the model—mannerisms—makes an impressive difference to the realism of the animats. Most game engines randomly play different animation cycles to add variation to the default animations (for instance, looking around while standing still). The mannerisms provide similar breaks from the standard animations, although they are driven by higher-level patterns in the animat's state. Mannerisms also offer unique insights into the emotional state of NPCs, which are much easier to identify than the implicit parameters for senses and actions.

The fuzzy sensations and emotions also imply that the changes in the manifestations are much smoother (for instance, loss of precision aiming). This difference is hard to notice when playing the game, but a spectator observing the animats closely can recognize the changes in crisp emotions—whereas the changes in fuzzy emotion do not stand out. Overall, from a spectator's point of view, the emotions implicitly seem less mechanical and somewhat more believable.

From a technical point of view, the fuzzy-state machines can be difficult to handle. Whether updated every time an event is received or on a regular basis, there's always the risk of the fuzzy state values becoming 0. Precautions need to be taken to prevent degenerate FFSMs, and in-depth experimentation must check their validity and adjust the transitions.

Although each concept benefited from tailored models (for instance, FFSA for sensations, HFSM for the moods), the system generally suffers from heterogeneity. A simpler design reusing only one type of finite-state technique could recapture the essence of the design, with extra simplicity. That said, the different components were extremely useful as proof of concepts, and demonstrated increasingly complex hierarchical architectures.