Group selection update

Group se­lec­tion might seem like an odd topic for a LessWrong post. Yet a google seach for “group se­lec­tion” site:less­wrong.com turns up 345 re­sults.

Just the power and gen­er­al­ity of the con­cept of evolu­tion is enough to jus­tify posts on it here. In ad­di­tion, the im­pact group se­lec­tion could have on the anal­y­sis of so­cial struc­ture, gov­ern­ment, poli­tics, and the ar­chi­tec­ture of self-mod­ify­ing ar­tifi­cial in­tel­li­gences is hard to over-es­ti­mate. David Sloan Wil­son wrote that “group se­lec­tion is ar­guably the sin­gle most im­por­tant con­cept for un­der­stand­ing the na­ture of poli­tics from an evolu­tion­ary per­spec­tive.” (You should read his com­plete ar­ti­cle here—it’s a much more thor­ough de­bunk­ing of the de­bunk­ing of group se­lec­tion than this post, al­though I’m not con­vinced his in­ter­pre­ta­tion of kin se­lec­tion is sen­si­ble.) And I will ar­gue that it has par­tic­u­lar rele­vance to the study of ra­tio­nal­ity.

Eliezer’s ear­lier post The Tragedy of Group Selec­tion­ism dis­misses group se­lec­tion, based on a math­e­mat­i­cal model by Henry Harpend­ing and Alan Rogers. That model is, how­ever, fatally flawed: It stud­ies the fix­a­tion of al­tru­is­tic vs. self­ish genes within groups of fixed size. The groups never go ex­tinct. But group se­lec­tion hap­pens when groups are se­lected against. The math used to ar­gue against group se­lec­tion as­sumes from the out­set that group se­lec­tion does not oc­cur. (This is also true of May­nard Smith’s fa­mous haystack model.)

(That post is still valuable; its main pur­pose is to ar­gue that math trumps wishes and aes­thet­ics. Em­piri­cal data, how­ever, trumps math­e­mat­i­cal mod­els.)

Nit­pick­ing di­gres­sion on definitions

“Group se­lec­tion” is one of those tricky phrases that doesn’t mean what it means. Deno­ta­tion­ally, group se­lec­tiond means se­lec­tion at the level of a group. Con­no­ta­tion­ally, though, group se­lec­tionc means se­lec­tion for al­tru­is­tic genes at the level of a group. This is be­cause, his­tor­i­cally, group se­lec­tion was posited to ex­plain ge­netic adap­ta­tions that are hard to ex­plain us­ing in­di­vi­d­ual se­lec­tion.

group se­lec­tionn, se­lec­tion at the group level for traits that are neu­tral or harm­ful at the level of the in­di­vi­d­ual, or that don’t even ex­ist within the in­di­vi­d­ual genome, should also be con­sid­ered. group se­lec­tionc is a sub­set of group se­lec­tionn is a sub­set of group se­lec­tiond. If group-level se­lec­tion oc­curs at all, then traits of the group that are not ge­netic traits, in­clud­ing cul­tural knowl­edge, must be con­sid­ered. That makes a huge differ­ence. Hu­man his­tory is full of group se­lec­tionn. Every time one group with bet­ter tech­nol­ogy or so­cial or­ga­ni­za­tion pushes an­other group off of its land, that’s at least group se­lec­tionn.

If you want to model evolu­tion thor­oughly, and se­lec­tion of groups oc­curs, then you need to model group se­lec­tiond to get your pre­dic­tions to match re­al­ity, even if group se­lec­tion oc­curs en­tirely as a re­sult of non-group se­lec­tionc ge­netic traits that provide ad­van­tages to in­di­vi­d­u­als. But peo­ple re­ject group se­lec­tiond on the ba­sis of ar­gu­ments lev­eled against group se­lec­tionc.

A case study of group se­lec­tionc: Nightshades

But I’m not back­ing off from say­ing that group se­lec­tion can ex­plain (some) al­tru­ism. Ed­ward Wil­son has been threat­en­ing for sev­eral years to write a book show­ing that group se­lec­tion is more im­por­tant than kin se­lec­tion for gen­er­at­ing al­tru­ism in ants. He doesn’t seem to have pub­lished the book, but you can read his ar­ti­cle about it. (Short ver­sion: Group se­lec­tion is es­pe­cially im­por­tant in ants be­cause ant colonies, which are small groups, en­gage in con­stant war­fare with each other.)

And this brings me to the rea­son for writ­ing this post now. Last week’s Science con­tained an ar­ti­cle by Emma Gold­berg et al. with the most clear-cut demon­stra­tion of group se­lec­tion that I have yet seen (sum­ma­rized here). It con­cerns flow­er­ing plants of the night­shade fam­ily (Solanaceae). They de­scend from plants that evolved self-in­com­pat­i­bil­ity (SI) about 90 mil­lion years ago. SI plants can’t pol­li­nate them­selves. This is a puz­zling trait. Sex­ual re­pro­duc­tion in it­self is puz­zling enough; but once a species is sex­ual, in­di­vi­d­ual se­lec­tion should drive out SI in fa­vor of self-com­pat­i­bil­ity (SC), the abil­ity to self-pol­li­nate. SC gives in­di­vi­d­u­als a great re­pro­duc­tive ad­van­tage—it means their offspring can con­tain 100% of their genes, rather than only 50%. The ad­van­tage given by SC is much greater than the sup­posed ad­van­tage of asex­ual over sex­ual re­pro­duc­tion: SC plants can both leave their own cloned offspring, and foist their genes onto the offspring of their neigh­bors at no ad­di­tional cost to them­selves. SC also makes sur­vival of their genes much more likely when a sin­gle plant is iso­lated far from oth­ers of its species; this, in turn, makes spread­ing over ge­o­graph­i­cal ar­eas eas­ier.

And yet, SI is a com­plex, multi-gene mechanism that evolved to pre­vent SC. Why did it evolve?

The au­thors looked at a phy­lo­ge­netic tree of 998 species of Solanaceae. In this tree, SI keeps de­volv­ing into SC. Be­ing an SC mu­tant in an SI species is the best of both wor­lds. You get to pol­li­nate your­self, and ex­ploit your al­tru­is­tic SI neigh­bors. When some mem­bers of an SI species go SC, we ex­pect those SC genes to even­tu­ally be­come fixed. And once a Solanaceae species loses SI and be­comes SC, it never re-evolves SI. This has been go­ing on for 36 mil­lion years. So why are so many species of Solanaceae still SI?

Let sI = spe­ci­a­tion rate for SI; eI = ex­tinc­tion rate for SI; rI = net rate of species di­ver­sifi­ca­tion = sI—eI. Like­wise, rC is the net rate of species di­ver­sifi­ca­tion for SC species. qIC is the rate of tran­si­tion from SI to SC. SI will be lost com­pletely if sI—eI = rI < rC + qIC = (sC—eC) + qIC.

The data shows that sC > sI, but eC >> eI, enough so that rI > rC + qIC. In English: Self-pol­li­na­tors spe­ci­ate and di­ver­sify more rapidly than SI species do, as we ex­pect be­cause SC pro­vides an in­di­vi­d­ual ad­van­tage. Once self-pol­li­na­tors evolve in an SI species, these ex­ploiters out-com­pete their al­tru­is­tic SI neigh­bors un­til the en­tire species be­comes SC. How­ever, SC species go ex­tinct more of­ten than SI species. This is thought to be be­cause SI makes a species less-likely to fix­ate dele­te­ri­ous genes (makes it more evolv­able, in other words). In­di­vi­d­ual se­lec­tion fa­vors SC; but species se­lec­tion fa­vors SI more than enough to bal­ance this out. No­tice that gene-based group se­lec­tion at the species level is math­e­mat­i­cally more difficult than group se­lec­tion at the tribal (or ant colony) level (ADDED: un­less there is ge­netic flow be­tween groups at the tribal/​colony level).

So let’s stop “ac­cus­ing” peo­ple of in­vok­ing group se­lec­tion. Group se­lec­tion is real.

Group rationality

Group se­lec­tion is es­pe­cially rele­vant to ra­tio­nal­ity be­cause, in an evolv­ing sys­tem, if we use the defi­ni­tion “Ra­tion­al­ists win,” “win­ning” ap­plies to the unit of se­lec­tion. In my painfully long post Only hu­mans can have hu­man val­ues, the sec­tions “In­stincts, al­gorithms, prefer­ences, and be­liefs are ar­tifi­cial cat­e­gories” and “Bias, heuris­tic, or prefer­ence?” ar­gue that the bound­ary be­tween an or­ganism’s bi­ases and val­ues is an ar­tifi­cial an­a­lytic dis­tinc­tion. Similarly, if group se­lec­tion hap­pens in peo­ple, then our dis­cus­sion of ra­tio­nal­ity and val­ues is overly fo­cused on the ra­tio­nal­ity and val­ues of in­di­vi­d­u­als, when group dy­nam­ics are part of what pro­duces ra­tio­nal (win­ning) group be­hav­ior.

Even if you still don’t be­lieve in group se­lec­tionc, you should ac­cept that group se­lec­tionn may al­low in­for­ma­tion to drift back and forth, in a fit­ness-neu­tral way, from be­ing stored in genomes, to be­ing cul­turally trans­mit­ted. And that makes it nec­es­sary, when talk­ing about ra­tio­nal­ity in a nor­ma­tive way, to con­sider the ra­tio­nal­ity of the group, and not just the ra­tio­nal­ity of its in­di­vi­d­u­als.

(This is re­lated to my un­pop­u­lar es­say Ra­tion­al­ists lose when oth­ers choose. When the unit of se­lec­tion is the group, rather than the in­di­vi­d­ual, the “choice” is made on the ba­sis of benefit to the group, rather than benefit to the in­di­vi­d­ual. This will pre­fer “ir­ra­tional” in­di­vi­d­u­als who ter­mi­nally (per­haps un­con­sciously) value benefits to the group, and not just benefits to them­selves, over “ra­tio­nal” self-in­ter­est.)

group se­lec­tionn makes the Pri­soner’s Dilemma and tragedies of the com­mons smaller prob­lems. But it raises a new prob­lem: Is the in­di­vi­d­ual the wrong place to put some of our col­lec­tive ra­tio­nal­ity? Since hu­mans have evolved in groups for a long time, the de­fault as­sump­tion is that at­tributes, such as our ra­tio­nal­ity, are already op­ti­mized for the unit of se­lec­tion.

Less gen­er­ally, if the group has already evolved to place some of our ra­tio­nal­ity into the group, what will hap­pen if we try to in­still it into the in­di­vi­d­u­als? Since group se­lec­tion is real, we can ex­pect to find situ­a­tions where mak­ing in­di­vi­d­u­als more ra­tio­nal up­sets the evolu­tion­ary equil­ibrium and makes the group win less. Un­der what cir­cum­stances will mak­ing in­di­vi­d­u­als more ra­tio­nal in­ter­act badly with group dy­nam­ics, and make our group less ra­tio­nal (= win less)? This will prob­a­bly oc­cur in cir­cum­stances in­volv­ing in­di­vi­d­ual al­tru­ism. But if the lo­cus of group ra­tio­nal­ity can drift from in­di­vi­d­ual genes to cul­tural knowl­edge, it may also oc­cur in situ­a­tions not in­volv­ing al­tru­ism.

Postscript: The long-term ne­ces­sity of war

If group se­lec­tion is partly re­spon­si­ble for hu­man al­tru­ism, this means that world peace may in­crease self­ish­ness. Kon­rad Lorenz made a sub­set of this claim in On Ag­gres­sion (1966): He claimed that the more effec­tive each in­di­vi­d­ual’s kil­ling tools are, the more nec­es­sary em­pa­thy is, to keep mem­bers of the group from kil­ling each other; then in­voked group se­lec­tion. (This seems to me to ap­ply a lot to ca­nines and not much to felines.) If group se­lec­tion works best with small groups, the switch from tribes to na­tion-states may have already be­gun this pro­cess. I do not, how­ever, no­tice markedly greater al­tru­ism in tribal groups than in na­tion-states.