research/652-windowed-perception-models

[jharwell]

Lerman2003 defines a MC amenable model for robots WITH memory, which could then
be used in discrete different equations to help model the average swarm state
for more complex controllers like DPO. So, new perception models should be
created:

• Add a new perception variant where robots can remember the last N objects they
  have seen, and do not ascribe a probability to their continued existence after
  they have left the robot's line of sight. Perform experiments to compare this
  to pheromone decay variant.

• Add a new perception variant where robots track all objects they have seen for
  the last N timesteps. Perform experiments to compare this to the pheromone
  decay variant.

• Add a new perception variant where robots can remember the last N sites they
  have visited and successfully picked up blocks from. If N=1, this is site
  fidelity (Lu2018,Just2017). Robots make stochastic decision each time they
  pick up a block on whether to overwrite their old remembered site with
  the current one, according to some criteria (density of nearby robots for
  example).

• Add a new perception variant where robots track object density via decaying
  poisson pheromone waypoints. Robots create a waypoint if they see more than R
  resources at/near a given site. Robots can exchange waypoints/access a shared
  pool of waypoints upon entering the nest, or not. From Hecker2015.

After defining the new perception models, they should be compared with the DPO
perception model to see where they are better/worse.