Behavioral walking uses behavioral or ``reactive planning'' techniques to navigate a simulated human or group of humans through a scene in order to get to a particular location without collisions [7]. The behavioral walking relies on tight couplings of simulated sensors to simulated effectors rather than on static plans based on a static environment. As a result, the behavioral walking works in real time and works effectively with moving obstacles (such as other agents), unexpected obstacles, or unexpected opportunities.
The behavioral walking is controlled through a network of attraction and avoidance behaviors connected to simulated object proximity sensors or simulated sonar arrays. The attract and avoid behaviors collectively determine the where each agent will step next - decisions about where to step next are made only at the beginning of each step.
Because the behavioral description is posed as a network of modules with floating point communication, low-level learning and numerical optimization techniques can be applied to aid the user in developing an adequate set of parameters for the behaviors. In [5], Becket uses gradient search and proposes to use a genetic optimization routine for finding optimal weights given a user-defined fitness function. This use of optimization addresses one of primary problems with designing behavioral or reactive systems - arbitration among competing behaviors to decide which gets control of the effectors rapidly becomes too complex for the agent designer. Also addressed in [5] is the sequencing of behavioral weights given to optimize total reward using a variety of unsupervised reinforcement learning techniques such as Q-learning and classifier systems.