The original alibi generation system cannot be precisely implemented as a subordinate of the LOD trader. With minor modifications, however, it can be, and gains desirable properties which were not available without the LOD trader. There are two potential forms for this integration, which will be described in turn.
5.8.1
Cells as entities
In this form, alibi generation as described in this chapter uses two separate feature graphs: One for cells, and one for agents.
The “cell” feature graph is the simpler of the two, and exists primarily to drive the rules for agent creation. It consists of one feature, with two LODs: visible and invisible. The “invisible” LOD, of course, has an extremely high audacity coefficient for unrealistic state. The cost of simulation at the “visible” LOD is based on the expected cost of simulation of newly created agents. The transition from invisible to visible also has significant resource costs, as it entails the creation of agents; however, the high audacity coefficient for unrealistic state means that, under all but the most extreme circumstances, this transition will occur whenever the cell region comes into view.
The “agent” feature graph is slightly more complex. It has one feature, with three LODs: alibi-less, alibi-ful, and nonexistent. As envisioned in the original design, transitions are only available from the first to the second, from the second to the third, and from the first to
the third LOD. in the alibi-less LOD, agents reaching portals whose other cell is visible do not immediately generate alibis, but transition to a new segment randomly based on a simple per-segment probability table. The alibi-less LOD has a small but significant audacity for unrealistic long-term behavior, causing its unrealism to increase over time while the agent is visible. Transitioning from the alibi-less to the alibi-ful LOD, of course, has a high CPU cost; the alibi-ful LOD has a slightly higher CPU and memory cost than the alibi-less LOD. The alibi-ful LOD has a zero audacity vector, reflecting its status as the “realistic” LOD. The transition from alibi-less or alibi-ful to nonexistent LOD has a high audacity coefficient for unrealistic state and a medium coeffient for fundamental discontinuity; coupled with the zero resource cost for the nonexistent LOD, this will cause un-memorable agents to be destroyed when out of view for a little while, but will prolong the lives of agents which have been the source of attention recently.
In addition, a transition may be added from the alibi-ful to the alibi-less LOD, with zero transition cost or penalty (the additional penalty coming from the destination level). This allows agents which were given alibis and subsequently unattended to discard those alibis for a small gain in resources.
5.8.2
Cells as triggers
In this form, two crowd simulations are used: the normal discrete agent simulation, and an overlaid flow-based simulation. Each cell will have a set of currently present discrete agents, and a real-valued continuous population.
At the beginning of the simulation, all cells have zero discrete population, and each has a continuous population equal to its average population. When a cell becomes visible following a period of invisibility (or is visible at the beginning of the simulation), as many agents are generated as the integral portion of the continuous population, with that population being correspondingly reduced. This transition only occurs at the moment when the cell
becomes visible.
Portals operate in a complementary manner. When both cells of a portal are invisible, it transfers continuous population in each direction with a flux proportional to the average flux, and to the incoming cell’s continuous population. When one cell of a portal is visible and the other is invisible, continuous population flows bidirectionally, but flux from the invisible cell to the visible cell does not increase the visible cell’s continuous population. Instead, a Poisson process generates discrete agents at the portal heading into the visible cell at the corresponding average rate.
Agents’ feature graph is similar to the previous form, with the added caveat that agents are not allowed to become nonexistent while in a visible cell (even if they themselves are not visible). When an agent does become nonexistent, its current cell’s continuous population is incremented by 1. Sufficiently memorable agents are thus allowed to continue operating in invisible cells, while less memorable agents are absorbed into the continuum.
5.8.3
Discussion
Both forms are reasonably straightforward to implement. The first form has a definite perfor- mance edge, allowing the instant destruction of swaths of agents when cells are sufficiently out of view. The second form, however, allows for arbitrary degrees of compromise between resource usage and realistic persistence. Moreover, it guarantees conservation of population, even in circumstances where the viewer’s actions change the normal flow of traffic.
Implementing the decision to create, destroy, and alibi-ize agents in terms of these benefit/cost tradeoffs removes some of the deterministic crispness of the original design, and compromises the elegant guarantee that alibi-less behavior can never be noticed. In particular, the benefit of giving an agent an alibi is calculated based on the duration of observation, rather than the number of inter-segment turns made. If desired, this latter
from the feature graph, and instead manually forcing it when the agent nears the end of the segment.
The benefit, however, is an effective solution to the concerns described in section 5.7. Agents are no longer destroyed needlessly or carelessly: the decision to destroy one is made only when the resources are needed, and the agents chosen for destruction are those the viewer is least likely to miss. Nor are alibis generated for agents almost immediately after creation, since the penalty for alibi-less simulation only becomes significant after a period of actual visibility (of the entity, not the cell). Finally, these LOD decisions may be coordinated with those of other simulation types, to provide better guarantees on total system performance.