Designing neural networks through neuroevolution

Stanley, Kenneth O.; Clune, Jeff; Lehman, Joel; Miikkulainen, Risto

doi:10.1038/s42256-018-0006-z

Cited by 566 publications

(336 citation statements)

References 108 publications

Supporting

Mentioning

332

Contrasting

Unclassified

Order By: Relevance

“…Thus, the probability for the transition

{\overset{x}{true}}_{i} \to {\overset{x}{true}}_{j}

is estimated at

{[{normalΓ}_{τ}]}_{i j} = \frac{M_{i j} false(τ false)}{M_{i}}

, where M ij (τ) is the number of transitions

{\overset{x}{true}}_{i} \to {\overset{x}{true}}_{j}

that occurred for the projected trajectory between t and t+τ within the time period

0 \leq t \leq t_{0} - τ

, and M i is the number of times state

{\bar{x}}_{i}

is visited within the same training time frame. Once this first estimation is at hand, the optimization of the transition operator requires deep learning as the projection of each MD‐generated microstate

x (t_{0} + n τ)

for n = 1 ,…, W and W satisfying

false(t_{0} + W τ false) \leq t_{f} < false[t_{0} + false(W + 1 false) τ false]

must be approximated as

{[{normalΓ}_{τ}]}^{n} π x false(t_{0} false) .

In other words, the parametrization of the transition operator in quotient space is optimized to minimize the loss function

L ({normalΓ}_{τ})

given by

\begin{matrix} true \\ right scriptL ({normalΓ}_{τ}) = W^{- 1} \sum_{n = 1}^{W} false∥ π x ((), t_{0} + n τ) - {false[Γ_{τ} false]}^{n} π x (t_{0}) {false∥}^{2} \end{matrix}

…”

Section: Theoretical Frameworkmentioning

confidence: 99%

“…We know this loss function is the correct one since

L ({normalΓ}_{τ}) = 0

if and only if the transition operator makes the previous diagram commutative. Once the optimal

{normalΓ}_{τ}^{*} = argmin

L ({normalΓ}_{τ})

has been obtained from stochastic steepest descent, the modulo‐basin projected trajectory can be propagated beyond MD‐accessible timescales. At this stage, the coarse‐grained modulo basin trajectory needs to be decoded back to the MD‐level atomistic description.…”

Section: Theoretical Frameworkmentioning

confidence: 99%

“…Once this first estimation is at hand, the optimization of the transition operator requires deep learning as the projection of each MD-generated microstate x(t 0 + n ) for n = 1,…,W and W satisfying (t 0 + W ) ≤ t f < [t 0 + (W + 1) ] must be approximated as [Γ ] n x(t 0 ). In other words, the parametrization of the transition operator in quotient space is optimized to minimize the loss function (Γ ) [16] given by…”

Section: Theoretical Frameworkmentioning

confidence: 99%

See 2 more Smart Citations

Deep Learning Unravels a Dynamic Hierarchy While Empowering Molecular Dynamics Simulations

Fernández

2020

Annalen der Physik

View full text Add to dashboard Cite

Molecular dynamics (MD) provide predictive understanding of the behavior of condensed matter. However, its true potential remains largely untested because relevant timescales are often inaccessible, limited portions of conformation space get sampled, and infrequent events are usually irreproducible. A culprit is the huge informational burden required to iterate integration steps. To address the problem, deep learning is applied to encode the dynamics into a shorthand embodiment retaining only essential topological features of the vector field that steers MD integration. The flow is simplified via an equivalence relation that identifies conformations within basins of attraction in potential energy and encodes the dynamics onto a modulo‐basin “quotient space” where fast motions are averaged out. The quotient space projection enables coverage of realistic timescales while unraveling the underlying dynamic hierarchy. Deep learning is exploited to propagate the simplified trajectory beyond MD‐accessible timescales and to reconstruct it at atomistic level. As shown, the quotient‐encoding‐propagating‐decoding scheme generates within a few hours protein folding pathways with experimentally verified outcomes. By contrast, MD computations covering comparable timespans would take over a hundred days on special‐purpose supercomputers. Thus, quotient space constitutes a model for hierarchical understanding of MD simulation while enabling access to realistic timescales.

show abstract

“…Thus, the probability for the transition

{\overset{x}{true}}_{i} \to {\overset{x}{true}}_{j}

is estimated at

{[{normalΓ}_{τ}]}_{i j} = \frac{M_{i j} false(τ false)}{M_{i}}

, where M ij (τ) is the number of transitions

{\overset{x}{true}}_{i} \to {\overset{x}{true}}_{j}

that occurred for the projected trajectory between t and t+τ within the time period

0 \leq t \leq t_{0} - τ

, and M i is the number of times state

{\bar{x}}_{i}

x (t_{0} + n τ)

for n = 1 ,…, W and W satisfying

false(t_{0} + W τ false) \leq t_{f} < false[t_{0} + false(W + 1 false) τ false]

must be approximated as

{[{normalΓ}_{τ}]}^{n} π x false(t_{0} false) .

In other words, the parametrization of the transition operator in quotient space is optimized to minimize the loss function

L ({normalΓ}_{τ})

given by

\begin{matrix} true \\ right scriptL ({normalΓ}_{τ}) = W^{- 1} \sum_{n = 1}^{W} false∥ π x ((), t_{0} + n τ) - {false[Γ_{τ} false]}^{n} π x (t_{0}) {false∥}^{2} \end{matrix}

…”

Section: Theoretical Frameworkmentioning

confidence: 99%

“…We know this loss function is the correct one since

L ({normalΓ}_{τ}) = 0

if and only if the transition operator makes the previous diagram commutative. Once the optimal

{normalΓ}_{τ}^{*} = argmin

L ({normalΓ}_{τ})

Section: Theoretical Frameworkmentioning

confidence: 99%

Section: Theoretical Frameworkmentioning

confidence: 99%

See 1 more Smart Citation

Deep Learning Unravels a Dynamic Hierarchy While Empowering Molecular Dynamics Simulations

Fernández

2020

Annalen der Physik

View full text Add to dashboard Cite

show abstract

“…For the Goal-ANN, we have no target value, since we do not know the "correct" goal. We therefore optimize this networks by using neuroevolution, a technique employing a population of neural networks, having the ones performing their task best become further adapted and specialized to the problem [14].…”

Section: Combining Deep Learning and Neuroevolutionmentioning

confidence: 99%

“…This Goal-ANN is trained using neuroevolution [14]: A population of neural networks compete for their ability to produce relevant goals. The best networks are randomly changed, by adding or removing nodes and connections, while worse networks are discarded.…”

Section: Adaptive Goals Guiding Action Selectionmentioning

confidence: 99%

Self-adapting Goals Allow Transfer of Predictive Models to New Tasks

Ellefsen

Tørresen

2019

Communications in Computer and Information Science

View full text Add to dashboard Cite

A long-standing challenge in Reinforcement Learning is enabling agents to learn a model of their environment which can be transferred to solve other problems in a world with the same underlying rules. One reason this is difficult is the challenge of learning accurate models of an environment. If such a model is inaccurate, the agent's plans and actions will likely be sub-optimal, and likely lead to the wrong outcomes. Recent progress in model-based reinforcement learning has improved the ability for agents to learn and use predictive models. In this paper, we extend a recent deep learning architecture which learns a predictive model of the environment that aims to predict only the value of a few key measurements, which are indicative of an agent's performance. Predicting only a few measurements rather than the entire future state of an environment makes it more feasible to learn a valuable predictive model. We extend this predictive model with a small, evolving neural network that suggests the best goals to pursue in the current state. We demonstrate that this allows the predictive model to transfer to new scenarios where goals are different, and that the adaptive goals can even adjust agent behavior on-line, changing its strategy to fit the current context.

show abstract

Artificial Intelligence and Advanced Materials

López

2023

Advanced Materials

View full text Add to dashboard Cite

Artificial intelligence (AI) is gaining strength, and materials science can both contribute to and profit from it. In a simultaneous progress race, new materials, systems, and processes can be devised and optimized thanks to machine learning (ML) techniques, and such progress can be turned into innovative computing platforms. Future materials scientists will profit from understanding how ML can boost the conception of advanced materials. This review covers aspects of computation from the fundamentals to directions taken and repercussions produced by computation to account for the origins, procedures, and applications of AI. ML and its methods are reviewed to provide basic knowledge of its implementation and its potential. The materials and systems used to implement AI with electric charges are finding serious competition from other information‐carrying and processing agents. The impact these techniques have on the inception of new advanced materials is so deep that a new paradigm is developing where implicit knowledge is being mined to conceive materials and systems for functions instead of finding applications to found materials. How far this trend can be carried is hard to fathom, as exemplified by the power to discover unheard of materials or physical laws buried in data.

show abstract

Designing neural networks through neuroevolution

Cited by 566 publications

References 108 publications

Deep Learning Unravels a Dynamic Hierarchy While Empowering Molecular Dynamics Simulations

Deep Learning Unravels a Dynamic Hierarchy While Empowering Molecular Dynamics Simulations

Self-adapting Goals Allow Transfer of Predictive Models to New Tasks

Artificial Intelligence and Advanced Materials

Contact Info

Product

Resources

About