This paper is totally awesome.
An evolutionary agent-based model was used in which agents had perceptual inputs and simple behaviors that were connected via a simple neural network. Network weights could evolve. Agents also had heritable "markers" which could also be perceived and used as inputs.
The model
- The world
- 2-D (30 x 30) grid.
- Each cell either contained a resource bundle or was empty. An empty cell could acquire a resource bundle with a certain probability per time step, and lose it when the resource is consumed by an agent.
- Agents are characterized by a set of receptors, a set of effectors, and a simple feed-forward neural net connecting them (no hidden layer).
- Agents' possible actions:
- Rest (do nothing)
- Eat (consume resource bundle if present)
- Divide (produce offspring)
- Move (go forward to a neighbor cell)
- Turn (turn left or right)
- Attack (attack another agent if present in same cell
- All actions expend resources, with rest costing the least and attack costing the most.
- If all internal resources are depleted, the agent dies.
- When an agent divides, an offspring is created in the same cell. The parent transfers half its resources to the offspring.
- Agent's field of vision: current cell, and cells forward, left, and right)
- Sensory inputs:
- Internal energy store
- Whether there are resources in the field of vision
- How many other agents in the field of vision.
- Each agent also has a "marker" that is coded by a vector of integers, and serves as an indicator of similarity.
- The Euclidean distance between an agent's markers and those of another agent in the cell (a potential subject for attack) is also a sensory input.
The total space of all possible strategies is huge. They group a lot together (e.g., they group 'rest', 'eat', and 'turn' together because they correspond to the absence of interaction between agents).
Results
- Without markers
- The game evolved into basic hawk-dove-bourgeois game.
- Hawk: Always attacks other agents nearby.
- Dove: Never attack, tries to move when attacked.
- Bourgeois: Say in the same cell and immediately attack any invader, but ignore agents in neighboring cells (unlike hawks).
- The dominant strategy is bourgeois, provided that the carrying capacity of a single cell is sufficient to support a sedentary agent (ie., the amount of resources in a bundle is sufficient for survival until the next bundle appears).
- Below this threshold (C1), the bourgeois strategy is imposible, and the population is divided between doves and hawks.
- The long-term population density increases linearly with carrying capacity until it reaches C1, and then becomes flat. This is because once the bourgeois takes over, each cell can be occupied by only a single agent. Even when resources can support more than one agent per cell, they fight until only one remains.
- With markers
- Agents can evolve the ability to detect phenotypic similarity.
- A different three strategies emerged:
- Cooperative dove: ignored out-group (large phenotypic distance) members, but left cells with in-group (small phenotypic distance) members.
- Raven: Left cells with in-group members but attacked out-group members when they detected them. Russian proverb: "a raven will not peck out the eye of another raven."
- Starling: Stay in the same cell with in-group members and (collectively) fight any out-group invader. Having to share resources meant that agents using this strategy had small stores of internal resources, but still had a good chance of defeating an invader with lots of resources because of their "mobbing" behavior advantage (hence the name).
- Emergence of the Starling strategy has effect on relationship between carrying capacity and long-term population density. For lower values of the carrying capacity, the curve in the "full" model follows that of the model without markers. Once it exceeds the threshold C2 = 2*C1, the density curve again rises. This is because, if the carrying capacity is not enough to support at least two agents in a cell (C < C2), the Starling strategy cannot invade. Instead, the dominant strategy is Raven, whose frequency increases linearly with C for C < C1.
- Once they can invade, Starlings do not drive Ravens to extinction. Instead, there are sustained oscillations between the numbers of Starlings, Ravens, and Doves.
- For low carrying capacity (C < C1), another new strategy emerged:
- Leave cell any time another agent (related or not) appeared in it.
- Starlings were also observed to develop the "desperado" effect. First predicted by Maynard Smith, if the internal resource of the agent falls below some level, the agent leaves the cell.
The strategy of "wolf," which would involve a cooperative attack, did not evolve, probably because agents lacked the appropriate effectors for travelling in groups in search of prey. In future, work, they play to add actions like "follow another agent" to see what could evolve.
No comments:
Post a Comment