genetic regulatory networks applied to neural networks

26
Genetic Regulatory Networks Applied to Neural Networks Bryan Adams MIT Computer Science and Artificial Intelligence Laboratory

Upload: yehudi

Post on 12-Jan-2016

49 views

Category:

Documents


0 download

DESCRIPTION

Genetic Regulatory Networks Applied to Neural Networks. Bryan Adams MIT Computer Science and Artificial Intelligence Laboratory. Outline. Motivation and Related Work System Overview and Results Conclusions. Motivation and Related Work System Overview and Results Conclusions. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Genetic Regulatory Networks Applied to Neural Networks

Genetic Regulatory Networks Applied to Neural Networks

Bryan AdamsMIT Computer Science and Artificial Intelligence

Laboratory

Page 2: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 2

• Motivation and Related Work

• System Overview and Results

• Conclusions

Outline

• Motivation and Related Work

• System Overview and Results

• Conclusions

Page 3: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 3

Motivation: June, 2004

Page 4: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 4

Motivation

• Robot controllers …– Robust

– Adaptive

– Complex behaviors

• Borrow from biology– Evolutionary Artificial

Neural Networks (ANNs)

– Genetic Regulatory Networks (GRNs)

Page 5: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 5

Motivation

Two similar robots (or cars) …

Slightly different morphologies

Page 6: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 6

• Husbands

– GasNets

• Zhou, Shen

– Bugs

• Stanley, Miikkulainen

– NEAT

Related Work: Evolutionary ANNs

Page 7: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 7

Related Work: GRNs

• Josh Bongard

– Artificial Ontogeny

• Peter Eggenberger

– Neural Retina

• Kumar

– GRN controller

Page 8: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 8

• Motivation and Related Work

• System Overview and Results

• Conclusions

Outline

• Motivation and Related Work

• System Overview and Results

• Conclusions

Page 9: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 9

System Overview: NEAT

• Direct, complete genetic encoding

• “Innovation numbers”– Very clever genetic

operators

– Speciation during evolution

• Theoretically minimal networks

Page 10: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 10

System Overview: GRN

• Repressive control

– Constitutively active

– Repressor shuts off

Pcnt

Prod

= Pcnt – (R Famt) ; >= 0

Prod

R

Page 11: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 11

System Overview: GRN

• Activator control

– Constitutively silent

– Activator causes expression

= A Famt ; <= Pcnt

0

Prod

Prod

A

Page 12: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 12

System Overview: Signals

• Decay according to first-order kineticst=1 = k t=0

• For n signals, half-lives are evenly spaced

Page 13: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 13

System Overview: NEAT-GRN

+

Environment

Page 14: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 14

System Overview: 36 NEAT Parameters

int n_links_avoid_chaining = 15;

int num_tries_insert_hid = 30;

float max_new_weight = 2.50f;

float max_big_weight = 10.0f;

float max_w_change = 2.50f;

bool allow_recurrent_links = false;

int num_tries_insert_link = 30;

float prob_reenable_during_xover = 0.25f;

float max_weight = 12.00f;

float min_weight =-12.00f;

float p_mutate_weights = 0.90f;

int min_size_age_prot = 10;

float old_links_frac = 0.20f;

float old_links_mul = 1.20f;

float p_severe_mut = 0.50f;

float p_severe_change = 0.70f;

float p_severe_new = 0.20f;

float p_normal_change = 0.50f;

float p_normal_new = 0.10f;

int min_size_for_elite = 5;

int max_elderly_amnesty = 15;

float failure_to_improve_penalty = 0.01f;

float good_parent_frac = 0.20f;

float p_mutate_only = 0.25f;

float p_inters_xover = 0.001f;

float upper_spec_frac = 0.22f;

float lower_spec_frac = 0.18f;

float dyn_spec_increment = 0.30f;

float c1 = 1.0f;

float c2 = 1.0f;

float c3 = 0.4f;

float delta_t = 3.0f;

float p_add_node = 0.03f;

float p_add_link = 0.30f;

float p_add_node = 0.001f;

float p_add_link = 0.05f;

Page 15: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 15

System Overview: 30 GRN Parameters

int n_signals = 4;

float max_half_life = 20;

float min_half_life = 2;

int production_steps = 50;

float signal_input_multiplier = 0.01f;

float lethal_fraction = 0.10f;

float p_take_both= 0.25f;

float p_add_copy_link = 0.15f;

float max_num_copies = 3;

float p_mutate_regl = 0.75f;

float p_add_regl = 0.00f;

float p_regl_severe_mut = 0.50f;

float p_regl_normal_chg = 0.50f;

float p_regl_normal_new = 0.10f;

float p_regl_severe_chg = 0.70f;

float p_regl_severe_new = 0.20f;

float c4 = 0.1f;

float p_no_prod = 0.50f;

float p_no_ra = 0.00f;

float p_neg_ctrl = 0.50f;

float famt_max_val = 0.30f;

float famt_max_incr = 0.02f;

float pcnt_max_val = 0.30f;

float pcnt_max_incr = 0.02f;

float p_change_rg = 0.00f;

float p_change_ra = 0.02f;

float p_change_pr = 0.04f;

float p_change_pc = 0.65f;

float p_change_fa = 1.00f;

float expression_amt = 0.00001f;

Page 16: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 16

Results: NEAT and XOR / NXOR

NEAT / XOR

Mean/SD Win gen 24.760 10.653Mean/SD Nodes 2.330 1.225Mean/SD Links 11.545 5.240

Results averaged over 200 runs; 100% solution success

NEAT / NXOR

Mean/SD Win gen 26.080 12.078Mean/SD Nodes 2.445 1.272Mean/SD Links 11.920 5.166

Page 17: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 17

Results: NEAT-GRN XOR / NXOR

Results averaged over 200 runs; 100% solution success

NEAT-GRN / XOR

Mean/SD Win gen 38.280 12.499Mean/SD Nodes 3.070 1.444Mean/SD Links 17.590 6.561

NEAT-GRN / NXOR

Mean/SD Win gen 37.945 14.586Mean/SD Nodes 3.040 1.612Mean/SD Links 17.880 7.608

Mean/SD Expr. Nodes 2.580 1.270Mean/SD Expr. Links 12.630 4.748

Mean/SD Expr. Nodes 2.545 1.299Mean/SD Expr. Links 12.720 5.091

Mean/SD Nodes 2.445 1.272Mean/SD Links 11.920 5.166

Mean/SD Nodes 2.330 1.225Mean/SD Links 11.545 5.240

Page 18: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 18

Results: NEAT-GRN Number of Signals

Results averaged over 200 runs; Same GRN parameters

0

10

20

30

40

50

60

1 2 3 4 5 6 7 8 9 10

Number Of Signals

Gen

era

tio

ns t

o W

inn

er

Page 19: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 19

Results: NEAT-GRN Number of Signals

Results averaged over 200 runs; Same GRN parameters

0

0.2

0.4

0.6

0.8

1

1.2

1 2 3 4 5

Percentage Postive Control

Su

cces

s P

erce

nta

ge

Page 20: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 20

Results: NEAT-GRN XOR & NXORResults averaged over 100 runs

35% solution success (max 250 gen)

Mean/SD Expr. Nodes 2.765 1.214Mean/SD Expr. Links 16.706 6.205

Mean/SD Nodes 2.330 1.225Mean/SD Links 11.545 5.240

NEAT-GRN XOR & NXOR

Mean/SD Win gen 164.294 47.157Mean/SD Nodes 4.353 1.713Mean/SD Links 32.824 11.984

NEAT-GRN / XOR

Mean/SD Win gen 38.280 12.499Mean/SD Nodes 3.070 1.444Mean/SD Links 17.590 6.561

Mean/SD Expr. Nodes 2.580 1.270Mean/SD Expr. Links 12.630 4.748

Mean/SD Expr. Nodes 2.545 1.299Mean/SD Expr. Links 12.720 5.091

Page 21: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 21

Results: XOR & NXOR networkbias_t [000] 0.000inpt_t [001] 0.000inpt_t [002] 0.000outp_t [003] 1.000hidn_t [004] 0.500 r{ + 1/0.029 0.054 }hidn_t [005] 0.500 r{ - 0/0.022 0.043 }hidn_t [007] 0.250 r{ - 0/0.030 0.049 }hidn_t [011] 0.125 r{ + 2/0.046 0.017 }hidn_t [013] 0.625 r{ - 1/0.012 0.096 }hidn_t [019] 0.313 r{ + 3/0.013 0.048 }link_t [000] [e] 0 3 5.88 r{ - 2/0.028 0.039 }link_t [001] [e] 2 3 -3.90 r{ - 2/0.007 0.037 }link_t [001] [e] 2 3 8.10 r{ + 0/0.021 0.005 }link_t [002] [e] 1 3 -6.04 r{ + 3/0.007 0.008 }link_t [003] [e] 2 4 2.70 r{ - 2/0.026 0.036 }link_t [004] [e] 4 3 11.65 r{ - 0/0.035 0.065 }link_t [005] [e] 1 5 -9.92 r{ + 1/0.007 0.040 }link_t [005] [e] 1 5 -10.55 r{ + 2/0.102 0.024 }link_t [005] [e] 1 5 -6.01 r{ + 1/0.021 0.014 }link_t [006] [e] 5 3 -8.85 r{ - 3/0.081 0.010 }link_t [007] [e] 1 4 6.69 r{ - 1/0.049 0.050 }link_t [008] [e] 0 4 -1.19 r{ - 2/0.088 0.005 }link_t [008] [e] 0 4 8.59 r{ - 2/0.045 0.019 }link_t [008] [e] 0 4 -4.49 r{ - 0/0.012 0.034 }link_t [009] [e] 0 5 1.68 r{ - 3/0.051 0.051 }link_t [010] [e] 2 5 3.76 r{ - 0/0.050 0.008 }link_t [010] [e] 2 5 3.41 r{ + 0/0.071 0.026 }link_t [010] [e] 2 5 -10.39 r{ - 3/0.054 0.067 }link_t [013] [e] 1 7 1.17 r{ + 3/0.097 0.022 }link_t [014] [e] 7 5 -12.00 r{ - 3/0.004 0.042 }link_t [017] [e] 2 7 7.27 r{ + 0/0.046 0.045 }link_t [018] [e] 7 4 -1.33 r{ - 2/0.023 0.040 }link_t [021] [e] 7 3 5.15 r{ - 2/0.006 0.059 }link_t [021] [e] 7 3 2.48 r{ + 0/0.031 0.030 }link_t [022] [e] 0 7 -3.24 r{ - 1/0.026 0.064 }link_t [022] [e] 0 7 7.14 r{ + 3/0.050 0.070 }link_t [025] [e] 1 11 7.35 r{ + 3/0.028 0.021 }link_t [026] [e] 11 7 5.03 r{ + 2/0.021 0.030 }link_t [029] [e] 7 13 1.27 r{ - 1/0.040 0.050 }link_t [030] [e] 13 3 -5.13 r{ - 0/0.020 0.041 }link_t [038] [e] 11 3 6.86 r{ - 2/0.032 0.043 }link_t [042] [e] 0 11 2.43 r{ - 1/0.011 0.027 }link_t [043] [e] 11 5 -1.64 r{ + 2/0.027 0.003 }link_t [044] [e] 2 11 -0.37 r{ + 3/0.013 0.023 }link_t [046] [e] 11 4 -8.51 r{ + 0/0.024 0.036 }link_t [053] [e] 11 19 -1.32 r{ - 3/0.058 0.074 }link_t [054] [e] 19 5 -7.22 r{ - 3/0.022 0.105 }link_t [055] [e] 19 4 -8.26 r{ + 3/0.076 0.031 }link_t [060] [e] 19 3 -4.63 r{ - 1/0.155 0.021 }link_t [064] [e] 1 19 -0.33 r{ + 2/0.048 0.016 }link_t [078] [e] 4 13 9.26 r{ + 3/0.078 0.029 }link_t [079] [e] 0 13 -2.86 r{ + 3/0.021 0.009 }link_t [080] [e] 2 13 2.33 r{ + 1/0.039 0.085 }link_t [080] [e] 2 13 -3.19 r{ - 1/0.086 0.061 }link_t [082] [e] 11 13 7.32 r{ + 3/0.034 0.042 }link_t [083] [e] 5 13 -2.39 r{ - 2/0.029 0.006 }link_t [085] [e] 1 13 3.85 r{ - 0/0.039 0.038 }link_t [093] [e] 2 19 -2.90 r{ - 1/0.020 0.021 }link_t [094] [e] 7 19 4.30 r{ + 1/0.001 0.056 }link_t [095] [e] 19 13 2.90 r{ - 2/0.046 0.025 }

Env1

Env0

Page 22: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 22

• Motivation and Related Work

• System Overview and Results

• Conclusions

Outline

• Motivation and Related Work

• System Overview and Results

• Conclusions

Page 23: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 23

Conclusions: Contributions

• Robust– A genome that can choose

between different expressions

• Adaptive– A controller where the env.

Feeds back to the GRN

• Complex behaviors – A genome that codes for

multiple behaviors

• A GRN model that features a variably-decoding phenotype

Page 24: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 24

Conclusions: Carsbias_t [000] 0.000inpt_t [001] 0.000inpt_t [002] 0.000outp_t [003] 1.000hidn_t [004] 0.500 r{ + 1/0.029 0.054 }hidn_t [005] 0.500 r{ - 0/0.022 0.043 }hidn_t [007] 0.250 r{ - 0/0.030 0.049 }hidn_t [011] 0.125 r{ + 2/0.046 0.017 }hidn_t [013] 0.625 r{ - 1/0.012 0.096 }hidn_t [019] 0.313 r{ + 3/0.013 0.048 }link_t [000] [e] 0 3 5.88 r{ - 2/0.028 0.039 }link_t [001] [e] 2 3 -3.90 r{ - 2/0.007 0.037 }link_t [001] [e] 2 3 8.10 r{ + 0/0.021 0.005 }link_t [002] [e] 1 3 -6.04 r{ + 3/0.007 0.008 }link_t [003] [e] 2 4 2.70 r{ - 2/0.026 0.036 }link_t [004] [e] 4 3 11.65 r{ - 0/0.035 0.065 }link_t [005] [e] 1 5 -9.92 r{ + 1/0.007 0.040 }link_t [005] [e] 1 5 -10.55 r{ + 2/0.102 0.024 }link_t [005] [e] 1 5 -6.01 r{ + 1/0.021 0.014 }link_t [006] [e] 5 3 -8.85 r{ - 3/0.081 0.010 }link_t [007] [e] 1 4 6.69 r{ - 1/0.049 0.050 }link_t [008] [e] 0 4 -1.19 r{ - 2/0.088 0.005 }link_t [008] [e] 0 4 8.59 r{ - 2/0.045 0.019 }link_t [008] [e] 0 4 -4.49 r{ - 0/0.012 0.034 }link_t [009] [e] 0 5 1.68 r{ - 3/0.051 0.051 }link_t [010] [e] 2 5 3.76 r{ - 0/0.050 0.008 }link_t [010] [e] 2 5 3.41 r{ + 0/0.071 0.026 }link_t [010] [e] 2 5 -10.39 r{ - 3/0.054 0.067 }link_t [013] [e] 1 7 1.17 r{ + 3/0.097 0.022 }link_t [014] [e] 7 5 -12.00 r{ - 3/0.004 0.042 }link_t [017] [e] 2 7 7.27 r{ + 0/0.046 0.045 }link_t [018] [e] 7 4 -1.33 r{ - 2/0.023 0.040 }link_t [021] [e] 7 3 5.15 r{ - 2/0.006 0.059 }link_t [021] [e] 7 3 2.48 r{ + 0/0.031 0.030 }link_t [022] [e] 0 7 -3.24 r{ - 1/0.026 0.064 }link_t [022] [e] 0 7 7.14 r{ + 3/0.050 0.070 }link_t [025] [e] 1 11 7.35 r{ + 3/0.028 0.021 }link_t [026] [e] 11 7 5.03 r{ + 2/0.021 0.030 }link_t [029] [e] 7 13 1.27 r{ - 1/0.040 0.050 }link_t [030] [e] 13 3 -5.13 r{ - 0/0.020 0.041 }link_t [038] [e] 11 3 6.86 r{ - 2/0.032 0.043 }link_t [042] [e] 0 11 2.43 r{ - 1/0.011 0.027 }link_t [043] [e] 11 5 -1.64 r{ + 2/0.027 0.003 }link_t [044] [e] 2 11 -0.37 r{ + 3/0.013 0.023 }link_t [046] [e] 11 4 -8.51 r{ + 0/0.024 0.036 }link_t [053] [e] 11 19 -1.32 r{ - 3/0.058 0.074 }link_t [054] [e] 19 5 -7.22 r{ - 3/0.022 0.105 }link_t [055] [e] 19 4 -8.26 r{ + 3/0.076 0.031 }link_t [060] [e] 19 3 -4.63 r{ - 1/0.155 0.021 }link_t [064] [e] 1 19 -0.33 r{ + 2/0.048 0.016 }link_t [078] [e] 4 13 9.26 r{ + 3/0.078 0.029 }link_t [079] [e] 0 13 -2.86 r{ + 3/0.021 0.009 }link_t [080] [e] 2 13 2.33 r{ + 1/0.039 0.085 }link_t [080] [e] 2 13 -3.19 r{ - 1/0.086 0.061 }link_t [082] [e] 11 13 7.32 r{ + 3/0.034 0.042 }link_t [083] [e] 5 13 -2.39 r{ - 2/0.029 0.006 }link_t [085] [e] 1 13 3.85 r{ - 0/0.039 0.038 }link_t [093] [e] 2 19 -2.90 r{ - 1/0.020 0.021 }link_t [094] [e] 7 19 4.30 r{ + 1/0.001 0.056 }link_t [095] [e] 19 13 2.90 r{ - 2/0.046 0.025 }

Env1

Env0

Page 25: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 25

Conclusions: Next Robots

Page 26: Genetic Regulatory Networks Applied to Neural Networks

October 15, 2004 Research Qualifying Exam 26

Long-term Objectives:Project Overview

I. Academica. Literature search / readingb. Qualifying examinationc. Thesis proposald. Doctoral dissertation

II. Robotic platforma. Design and fabricationb. Robot chassis and motor systemc. Sensors and camerasd. Firmware and drivers

III. Softwarea. Artificial brain modules:

i. NEATer with GRNii. NEATer with developmentiii. NEATer with topologyiv. Synthetic Brains (integrated)

b. Simulation and evolution:i. Simulated arm and motors

ii. Simulated sensors iii. Evolutionary algorithm

An outline of the work to be done between now and October ‘05