Altruism in the Evolution of Communication

Altruism in the Evolution of Communication

David H. Ackley Michael L. Littman From Artificial Life III, 1994

Speaker: Omri Tal, otb@math.tau.ac.il Artificial Life seminar, Tel Aviv Uni, spring ’98 , Dr. Eitan Ruppin Introduction

Existing models for the evolution of communication directly reward speakers for improved behavior on the part of the listeners.
In these models fitness is effective communication.
In the new model effective communication evolves, even though ‘speaking truthfully’ provides no tangible benefit to the speaker.
Communication range approximately corresponds with breeding range, allowing kin selection to encourage the emergence of communication.
This also creates opportunities for exploitation by ‘information parasites’.
Darwinian evolution is about never-ending competition, but cooperation among individuals also plays a big role.
We focus on initially arbitrary signals.
We wish to eliminate the possibility of ‘mimetic semantics’.
Focus on the evolving of speaking and understanding simultaneously.
A model that rewards both parties for conversation does not develop lying.
A model in which communication causes males to find females more rapidly, did not develop speciation.

Population Levels

Three organizational levels of the population.
Individual level: A Genotype that codes the brain, which produces the Phenotype.
The Individual is short-sighted but has good hearing and speaking abilities.
Local level: Communication and reproduction between 8 individuals, this is the sub-population.
All 8 individuals get the same stimuli arrangement. This is the key to sharing information.
Global level: 128 X 128 grid of sub-populations, migration and reproduction.

A Day in the Life

A round of behavioral scoring is a day: 36 repetitions of independent trials.
Each trial is 3 moves. After each trial the score is adjusted as a function of the food, predator in sight, it’s location, and how much it attempted to move.
Unless the individuals are actually communicating, the resulting scores can be maximum –12. At end of day we have reproduction and death.

The genome

The genome is a 550 bit array, that defines a wiring diagram and initial conditions for the brain.
The genome is sequentially read to build the brain.
The brain is a synchronously updated neural network containing 32 linear threshold units: 12 sensor units, 8 effector units, 12 hidden units and 1 true unit.
All outputs are 0/1. 1, if weighted sums are greater than 0, otherwise 0. Two passes are performed each step.

Reproduction

We use crossover
No mutation is used – found to be too destructive.
Separate parallel simulations revealed poor behavioral scores.
Need to consider migration.
Two mechanisms: wind and festivals.
Wind: random selection of the immigrant and direction.
Festival: held in a quad and the child immigrates. Quads are shifted so immigration could be really global.
Windy festival days were also studied.
The simulation times were measured in weeks.
Festival-only was the most successful.

Wind-Only

After 13000 days it seemed to stabilize.
Maximum scores jumped to 42 (communication) very fast.
Average stabilized to a bit below –12.
Good communication strategies could score more than 42.
An organism type is the score it obtains in a sub-population of clones.
Not moving at all is the most probable behavior score for a randomly generated individual. They score –696.
-56’s started to take over. They run to one end, and then back if predator is seen.
-12’s also appeared.
26’s appeared, but were almost became extinct by –12’s, that have discovered a way to mislead them.
Altruism has a hard time stabilizing here. It seems that wind helps ‘cheaters’ invade.

Festivals and Wind

Festival is a score-sensitive migration.
Festival every other day, and wind every 10^th day.
Stabilized after about 14,000 days.
58’s appeared but soon squashed by disruptive –12’s.
Communicators spread farther and faster here, but still unstable.

Strictly Festivals

Stabilized after ~ 100,000 days.
Communicating organisms begin to appear after ~ 5000 days.
142’s are the best communicators found.
Under this model, it is generally not possible for one disruptive individual to enter and exterminate a high-scoring sub-population.

Epilogue

Communication based on arbitrary signals can evolve and stabilize even when it provides no benefit for the individual speaker.
The principal feature of the model is the partial alignment of communication and reproduction domains.
Full alignment would result only in mediocre fitness scores, as motivation for real communication is lacking.
Small correlation between communication and reproduction domains would also fail, since good communicators would not be given a chance to 'teach' other sub-populations.

------------------------------------------------------------------------------------------------------------------------------------------------------- My Remarks and Impressions Support

It is exciting to see that altruistic behavior and truthful communication tends to succeed in raising the overall fitness of a sub-population.
Also illuminating is the prospect that lying organisms tend to fail in the long run, although they are initially successful in invasion to a different sub-population.

Objections

This is similar to the model in which communication allows males to find females more rapidly. Here communication raises the probability of being the parents in the sub-population.
In real life speech is a learned ability. Only the capacity for complex semantics evolves.
The motivation for introducing Festivals is unclear. The parents may be from different societies, speaking different languages. The child may be born in another society.
Only 1 or 2 runs of each model were performed.

Improvements

It could be interesting to change the immigration: Both parents immigrate after mating, or one parent and it’s child, so that they could keep the dialect developed and teach others in a more primitive society.
I suggest death of the least-fit individual instead of a random selection.