on selection pressure
Locked
 
|
Here are a couple of plots from a large constrained optimization problem I've been running.
In the first case, I apply selection pressure: If a solution is not in the top 200 performers, it dies. In the second case, the population can continue to grow without concern for its performance. This is a 5900-dimensional pseudo-boolean problem and the best-known solution is around 2.61e+08. Note the low end of the y axis is not close to this. In both cases, aggressive efforts are made to diversify the population and in both cases every shown solution is unique (even though their energies can collide). In this case, I would argue that selection pressure has accomplished nothing -- conservatism doesn't work if the goal is to create the most fit individuals. The mean moves, if you care about that. But the very best solutions are nearly the same, and neither have come close to the optimal. Marcus ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
diversity.png (8K)
|
Are there computational (or otherwise not shown) costs to the members that continue in the free case but are pruned in the selection case?
On 1/2/19 7:44 AM, Marcus Daniels wrote: > Here are a couple of plots from a large constrained optimization problem I've been running. > In the first case, I apply selection pressure: If a solution is not in the top 200 performers, it dies. > In the second case, the population can continue to grow without concern for its performance. > This is a 5900-dimensional pseudo-boolean problem and the best-known solution is around 2.61e+08. Note the low end of the y axis is not close to this. In both cases, aggressive efforts are made to diversify the population and in both cases every shown solution is unique (even though their energies can collide). > > In this case, I would argue that selection pressure has accomplished nothing -- conservatism doesn't work if the goal is to create the most fit individuals. The mean moves, if you care about that. But the very best solutions are nearly the same, and neither have come close to the optimal. -- ∄ uǝʃƃ ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove
uǝʃƃ ⊥ glen
|
|
|
Some memory, and the ongoing recombination and optimization of less fit (high energy) individuals which tend to create other less fit individuals.
In this optimization system there are numerous methods that are used to create fit individuals, but the ones that create the very best individuals do not arise from recombination + selection pressure. Mixing two distinct (large Hamming distance) globally constraint-satisfying solutions tends to create a non-constraint satisfying solutions. It is only once the two parents are very similar (e.g. same species) that such a recombination will even work, but by then it doesn't do all that much. Computationally, it easier to try more approaches and maintain a large population than it is accelerate the algorithms that are most effective. (For the former, just add more cores.) On 1/2/19, 8:57 AM, "Friam on behalf of ∄ uǝʃƃ" <[hidden email] on behalf of [hidden email]> wrote: Are there computational (or otherwise not shown) costs to the members that continue in the free case but are pruned in the selection case? On 1/2/19 7:44 AM, Marcus Daniels wrote: > Here are a couple of plots from a large constrained optimization problem I've been running. > In the first case, I apply selection pressure: If a solution is not in the top 200 performers, it dies. > In the second case, the population can continue to grow without concern for its performance. > This is a 5900-dimensional pseudo-boolean problem and the best-known solution is around 2.61e+08. Note the low end of the y axis is not close to this. In both cases, aggressive efforts are made to diversify the population and in both cases every shown solution is unique (even though their energies can collide). > > In this case, I would argue that selection pressure has accomplished nothing -- conservatism doesn't work if the goal is to create the most fit individuals. The mean moves, if you care about that. But the very best solutions are nearly the same, and neither have come close to the optimal. -- ∄ uǝʃƃ ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
|
|
Very cool, Marcus! Did you interact with Ken Stanley (https://scholar.google.se/citations?user=6Q6oO1MAAAAJ&hl=en&oi=ao) when he was at SFI a couple years back? Ken's research would support your observations on the importance on the pressure to maintain novelty/diversity in evolutionary algorithms vs the focus on the objective function. In particular this paper: Also, Ken's homepage: http://www.cs.ucf.edu/~kstanley/ with more popular book links and Santa Fe Radio Cafe Interviews. BTW, in the late 90's I was working a bit on evolving weights and topologies of neural networks and was very inspired by Ken's advisor, Risto Miikkulainen, and his team at UT Austin: _______________________________________________________________________ [hidden email] CEO, Simtable http://www.simtable.com 1600 Lena St #D1, Santa Fe, NM 87505 office: (505)995-0206 mobile: (505)577-5828 twitter: @simtable On Wed, Jan 2, 2019 at 9:11 AM Marcus Daniels <[hidden email]> wrote: Some memory, and the ongoing recombination and optimization of less fit (high energy) individuals which tend to create other less fit individuals. ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
|
|
Hi Stephen, Thanks for the paper. I have some colleagues that study deceptive energy landscapes but it is a different literature. [I do like trying to figure out “How the hell did that work?!” more than a workman-like construction project. Maybe up to a point when the experiments and reverse engineering just get to be too hard (biology).] Marcus From: Friam <[hidden email]> on behalf of Stephen Guerin <[hidden email]> Very cool, Marcus! Did you interact with Ken Stanley (https://scholar.google.se/citations?user=6Q6oO1MAAAAJ&hl=en&oi=ao) when he was at SFI a couple years back?
Ken's research would support your observations on the importance on the pressure to maintain novelty/diversity in evolutionary algorithms vs the focus on the objective function. In particular this paper: Also, Ken's homepage: BTW, in the late 90's I was working a bit on evolving weights and topologies of neural networks and was very inspired by Ken's advisor, Risto Miikkulainen, and his team at UT Austin:
_______________________________________________________________________ CEO, Simtable http://www.simtable.com 1600 Lena St #D1, Santa Fe, NM 87505 office: (505)995-0206 mobile: (505)577-5828 twitter: @simtable On Wed, Jan 2, 2019 at 9:11 AM Marcus Daniels <[hidden email]> wrote:
============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
|
|
Here's Ken's TV interview in Santa Fe which provides a nice review of the findings in his book. _______________________________________________________________________ [hidden email] CEO, Simtable http://www.simtable.com 1600 Lena St #D1, Santa Fe, NM 87505 office: (505)995-0206 mobile: (505)577-5828 twitter: @simtable On Wed, Jan 2, 2019 at 9:49 AM Marcus Daniels <[hidden email]> wrote:
============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
|
|
The connections are imperfect, but the world is better with people like Ken that are willing to explain them. If you want to execute perfect arithmetic, use a computer. From: Friam <[hidden email]> on behalf of Stephen Guerin <[hidden email]> Here's Ken's TV interview in Santa Fe which provides a nice review of the findings in his book.
_______________________________________________________________________ CEO, Simtable http://www.simtable.com 1600 Lena St #D1, Santa Fe, NM 87505 office: (505)995-0206 mobile: (505)577-5828 twitter: @simtable On Wed, Jan 2, 2019 at 9:49 AM Marcus Daniels <[hidden email]> wrote:
============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
|
|
In reply to this post by Marcus G. Daniels
The reason I asked was your statement "selection pressure has accomplished nothing". What I would be looking for is a more comprehensive description of the solution space showing selection as selecting a *subset* of properties/dimensions of the space. So, while selection may not have pushed the population very effectively in that subset, it may have changed the population's character in the whole space. So, a better statement would be "selection pressure didn't accomplish what it was intended to accomplish." Right?
> On 1/2/19 7:44 AM, Marcus Daniels wrote: > > In this case, I would argue that selection pressure has accomplished nothing -- conservatism doesn't work if the goal is to create the most fit individuals. The mean moves, if you care about that. But the very best solutions are nearly the same, and neither have come close to the optimal. -- ☣ uǝlƃ ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove
uǝʃƃ ⊥ glen
|
|
|
Sure, you can see that selection pressure samples more intensively in the lower quantiles of energy, just not close to the ground state. The goal I had was to find the ground state, not just to give the appearance of one. Stepping back, I claim what conservatives want is to give the appearance of fitness (and justify the current social order), not to achieve greatness for its own sake.
On 1/2/19, 12:21 PM, "Friam on behalf of uǝlƃ ☣" <[hidden email] on behalf of [hidden email]> wrote: The reason I asked was your statement "selection pressure has accomplished nothing". What I would be looking for is a more comprehensive description of the solution space showing selection as selecting a *subset* of properties/dimensions of the space. So, while selection may not have pushed the population very effectively in that subset, it may have changed the population's character in the whole space. So, a better statement would be "selection pressure didn't accomplish what it was intended to accomplish." Right? > On 1/2/19 7:44 AM, Marcus Daniels wrote: > > In this case, I would argue that selection pressure has accomplished nothing -- conservatism doesn't work if the goal is to create the most fit individuals. The mean moves, if you care about that. But the very best solutions are nearly the same, and neither have come close to the optimal. -- ☣ uǝlƃ ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com archives back to 2003: http://friam.471366.n2.nabble.com/ FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove |
| Free forum by Nabble | Edit this page |