The evolution and stability of cooperative traits [Protected Link] The evolution and stability of cooperative traits


  author         = {Sandip Sen and Partha Sarathi Dutta},
  title          = {The evolution and stability of cooperative traits},
  year           = {2002},
  review-dates   = {2004-07-31},
  value          = {aa},
  address        = {Bologna, Italy},
  booktitle      = {Proceedings of AAMAS '02},
  month          = {July},
  harcopy        = {yes},
  key            = {sen-2002a}


More realistic because of dynamic This papers explores how dynamically changing behaviors, rather than fixed behaviors changes the outcome in the evolution of a system of agents, with the population of behaviors changing over time (agents adapt, not just generations/generational mixes). Also here Sen uses more than two behavior types and adds potentially disruptive mixes into the populations (such as philantropic agents) into

Uses sigmoid function to determine likelihood of cooperation.

Eq #1 here

What is the likelihood (realism) of actually knowing /{B_ij}? for implementing "believing reciprocative agents". Or / and B_kj>0 B_ji for earned-trust agents?

!! The curves of Figure 2 show just how important effective observation and/or communication of cooperativeness of other agents is. The believing agent populations root out the selfish very quickly compared to the merely reciprocative. [Here's an opportunity of observation and communication networks to come into play, and see how the curve degrades over different networks of information sharing.]

From Figure 4, it appears a hybrid of believing in initial stages, and earned trust in later stages might be dominant over the other two... but only in the sense that they would maximize population a stable level, not actually wipe the others out.

The information processing domain studies weren't able to fit fully in here, though results were described as "qualitatively similar". The key distinction (not clear to me from earlier description, was that the package delivery domain was such that help-giving behavior influenced the likelihood of interacting with other agents (can't ask for help "on-the-road").

I'm not sure I grasp the following in Section 4. " can appear to be better than the rest either by doing well itself by ruining others."

The lying agents are hurting those (lying about) those that help them the most? This is malicious, but is it clever? Why not lie or exagerate only about those who are nice to others but not to oneself? Wouldn't one want to identify and "kill" those agents that are discriminating? When collobaratively lying, why not merely exaggerate helpfulness of other selfish agents? Do selfish collaborators also help each other under normal rules of conduct?

What are tau (steepness, approaching step function) and Beta (inclination to cooperate) constants? 0.75 and 1.0, given Figure 1. Does there valuation change experiments? How many values were tried... were they borrowed from fixed behavior values? How established? Is C^k_avg include all the help provided as part of C^kl_ij efforts, or just its own costs?

"The cost of executing a task can be reduced or eliminated if help is obtained from another agent." Although the discussion of the "possibility for cooperation" would indicate the costs should be less for the helper than the help requester, a selfish agent in the experiments wil ask for help even if the cost of doing it itself is *less* than the cost of the other doing it (which is what selfish agents would do!)

In Figure 3, presumably philanthropic agents are started at 30% with the others at 35% because it a) won't matter, b) makes graph a little easier to read, and c) in almost any real-life scenario, this very high amount of philanthropic agents significantly exceeds the amount likely to be found.

I'm not sure we explicitly compare/constract the fixed behavioral findings versus the dynamic ones. This is probably obvious by reading previous papers of Sen et al.

Experiment setup:



Discussion of probability function vs. distribution and Figure 1 seems a little blurry. - Probabilistic reciprocity scheme... whether to honor request for help from another agent. + Near-optimal homogenous + Outperfs exploiters in heterogenous - composition of group in term of agent behaviors is fixed (real-life agents will change their behavior based on observed performances of different behavioral traits) - Found ecological niches for variants of exploitative selfish agents and robust reciprocative agents [Note to self: This is more applicable to older area of interest/approach than emerging approach?] Again emphasizes "open, heterogenous environments"

Key Factors

How placed in context (other work): axelrod-1984a - Sen is critical of IPD-like games and their lessons for real-life application due to most underlying assumptions that motivate its use are viloated. (prior work). Prior work of Sen sen-1996a, sen-2000a. Also cooperation level - goldman-1994a, biological altruism - trivers-1972a, IPD - rappaport-1989a, ESS - day-1998a.

Not modeling altruistic behavior in animals or humans. "Not addressing issues raised by social science or experimental economics literature...".

Problem Addressed: Dynamism of behavior, advanced selfish collaboration (group selfishness, as it were). Do findings from fixed behavior studies hold under dynamic behavior modification by adaptive agents?

Main Claim and Evidence:

Assumptions: "We assume that typical real-world environments abound in cooperation possibilities: situations where one agent can help another agent by sharing work such that the helping cost of the helper is less than the cost saving of the helped agent." [Note: In this definition, only the cost savings to the system are noted, not any net benefit to the individual offering help.]

Not an assumption: fixed behavior. Assumption of fixed behaviors is a problem in prior/other work, too restrictive. Give agents choice of behaviors.... evolutionary process, with dynamically changing composition of

Next steps: "Analytically capture the dynamics of the evolution of agent population. Given a particular group composition and tasks per evaluation period, we plan to analyze and predict the behavioral composition of the group over time." - Investigate alternate schemes for adopting new behaviors (e.g., neighborhood-based schemes (rather than global sampling) - Look for group/coalition formation evolution amongst agents with mutually complementary behavior.

Remaining open questions:


Originality is excellent.
Contribution/Significance is outstanding.
Quality of organization excellent.
Quality of writing is excellent.