Skip to content

model.md

Latest commit

 

History

History
50 lines (33 loc) · 3.4 KB

model.md

File metadata and controls

50 lines (33 loc) · 3.4 KB
Raw

Model overview

This repository implements a small active-inference-inspired MDP simulation of ant-colony foraging. The paper describes the MDP as the decision process of an individual simulated ant; colony-level behavior emerges from many ants interacting through the shared environment and its pheromone field.

Environment

Env in ants.py maintains a 40 × 40 grid (config.py). Every grid cell has one of ten discrete pheromone observations. The initial observation is level 0 everywhere. Returning ants increase the observation level of each cell on their path; Env.decay() probabilistically decreases nonzero levels.

The nest starts at (20, 30). The food region is centered at (40, 5) with a nominal size of (10, 10). Movement is constrained to the interior of the grid and checked against the T-maze walls. An ant explores until it reaches the food region and the nest-return transition is sampled; it then follows its stored trajectory backward, depositing pheromone and eventually returning to exploration.

Ant and MDP

Each Ant owns an MDP and stores its current position and forward trajectory. The action map has nine one-step actions: the eight neighboring displacements plus staying in place. An ant's local likelihood matrix is rebuilt from the nine cells in its focal area at each forward step.

The MDP contains:

  • A: observation likelihoods for the nine local hidden states;
  • B: action-conditioned state transitions;
  • C: preferred observations;
  • sQ: inferred state beliefs; and
  • uQ: action probabilities.

MDP.step() performs state inference using the previous action and then scores actions using the expected-free-energy calculation implemented in the source. It samples the next action from the resulting softmax distribution. The code also applies priors against staying in place and immediately reversing the previous action.

Experiments

The three executable scripts define the experiments used by this repository:

  • figure_1.py runs one 70-ant, 2,000-step experiment with a food-location switch and writes the primary animation, snapshots, summary, and completed path plots.
  • figure_2.py repeats five runs for colonies of 10, 30, 50, and 70 ants using a gradual observation preference.
  • figure_3.py repeats five runs for colonies of 10, 30, 50, and 70 ants using a strict preference that assigns increasing value to pheromone levels 5–9. Its preference vector is transposed to the column shape expected by MDP.

Each script uses save=True and writes generated artifacts below imgs/. The defaults are computational experiments, not a packaged command-line interface.

Main parameters

config.py is the single source for the simulation dimensions and environment constants. Important values include:

Parameter Current value Meaning
GRID [40, 40] Environment dimensions
FOCAL_AREA [3, 3] Local observation window
NUM_OBSERVATIONS 10 Discrete pheromone levels
NUM_ACTIONS 9 One-step actions
ADD_ANT_EVERY 50 Steps between additions in main()
DECAY_FACTOR 0.01 Per-cell probability of one-level decay
NEST_FACTOR 0.1 Probability of switching to return mode after food contact
MAX_LEN 500 Defined trajectory limit (not currently enforced)

The implementation is stochastic and does not set a random seed. Individual runs therefore need not produce identical outputs.