Genetic Regulatory Networks Applied to Neural Networks Bryan
Genetic Regulatory Networks Applied to Neural Networks Bryan Adams MIT Computer Science and Artificial Intelligence Laboratory October 15, 2004 Research Qualifying Exam
Outline • Motivation and Related Work • System Overview and Results • Conclusions October 15, 2004 Research Qualifying Exam 2
Motivation: June, 2004 October 15, 2004 Research Qualifying Exam 3
Motivation • Robot controllers … – Robust – Adaptive – Complex behaviors • Borrow from biology – Evolutionary Artificial Neural Networks (ANNs) – Genetic Regulatory Networks (GRNs) October 15, 2004 Research Qualifying Exam 4
Motivation Two similar robots (or cars) … Slightly different morphologies October 15, 2004 Research Qualifying Exam 5
Related Work: Evolutionary ANNs • Stanley, Miikkulainen – NEAT • Husbands – Gas. Nets • Zhou, Shen – Bugs October 15, 2004 Research Qualifying Exam 6
Related Work: GRNs • Kumar – GRN controller • Peter Eggenberger – Neural Retina • Josh Bongard – Artificial Ontogeny October 15, 2004 Research Qualifying Exam 7
Outline • Motivation and Related Work • System Overview and Results • Conclusions October 15, 2004 Research Qualifying Exam 8
System Overview: NEAT • Direct, complete genetic encoding • “Innovation numbers” – Very clever genetic operators – Speciation during evolution • Theoretically minimal networks October 15, 2004 Research Qualifying Exam 9
System Overview: GRN • Repressive control Pcnt – Constitutively active – Repressor shuts off Prod R Prod = Pcnt – (R Famt) ; >= 0 October 15, 2004 Research Qualifying Exam 10
System Overview: GRN • Activator control 0 Prod – Constitutively silent A – Activator causes expression Prod = A Famt ; <= Pcnt October 15, 2004 Research Qualifying Exam 11
System Overview: Signals • Decay according to first-order kinetics t=1 = k t=0 • For n signals, half-lives are evenly spaced October 15, 2004 Research Qualifying Exam 12
System Overview: NEAT-GRN Environment + October 15, 2004 Research Qualifying Exam 13
System Overview: 36 NEAT Parameters int n_links_avoid_chaining = 15; float p_normal_new = 0. 10 f; int num_tries_insert_hid = 30; int min_size_for_elite = 5; float max_new_weight = 2. 50 f; int max_elderly_amnesty = 15; float max_big_weight = 10. 0 f; float failure_to_improve_penalty = 0. 01 f; float max_w_change = 2. 50 f; float good_parent_frac = 0. 20 f; bool allow_recurrent_links = false; float p_mutate_only = 0. 25 f; int num_tries_insert_link = 30; float p_inters_xover = 0. 001 f; float prob_reenable_during_xover = 0. 25 f; float upper_spec_frac = 0. 22 f; float max_weight = 12. 00 f; float lower_spec_frac = 0. 18 f; float min_weight =-12. 00 f; float dyn_spec_increment = 0. 30 f; float p_mutate_weights = 0. 90 f; float c 1 = 1. 0 f; int min_size_age_prot = 10; float c 2 = 1. 0 f; float old_links_frac = 0. 20 f; float c 3 = 0. 4 f; float old_links_mul = 1. 20 f; float delta_t = 3. 0 f; float p_severe_mut = 0. 50 f; float p_add_node = 0. 03 f; float p_severe_change = 0. 70 f; float p_add_link = 0. 30 f; float p_severe_new = 0. 20 f; float p_add_node = 0. 001 f; float p_normal_change = 0. 50 f; float p_add_link = 0. 05 f; October 15, 2004 Research Qualifying Exam 14
System Overview: 30 GRN Parameters int n_signals = 4; float p_regl_severe_new = 0. 20 f; float max_half_life = 20; float c 4 = 0. 1 f; float min_half_life = 2; float p_no_prod = 0. 50 f; int production_steps = 50; float p_no_ra = 0. 00 f; float signal_input_multiplier = 0. 01 f; float p_neg_ctrl = 0. 50 f; float lethal_fraction = 0. 10 f; float famt_max_val = 0. 30 f; float p_take_both= 0. 25 f; float famt_max_incr = 0. 02 f; float p_add_copy_link = 0. 15 f; float pcnt_max_val = 0. 30 f; float max_num_copies = 3; float pcnt_max_incr = 0. 02 f; float p_mutate_regl = 0. 75 f; float p_change_rg = 0. 00 f; float p_add_regl = 0. 00 f; float p_change_ra = 0. 02 f; float p_regl_severe_mut = 0. 50 f; float p_change_pr = 0. 04 f; float p_regl_normal_chg = 0. 50 f; float p_change_pc = 0. 65 f; float p_regl_normal_new = 0. 10 f; float p_change_fa = 1. 00 f; float p_regl_severe_chg = 0. 70 f; float expression_amt = 0. 00001 f; October 15, 2004 Research Qualifying Exam 15
Results: NEAT and XOR / NXOR Results averaged over 200 runs; 100% solution success October 15, 2004 Research Qualifying Exam 16
Results: NEAT-GRN XOR / NXOR Results averaged over 200 runs; 100% solution success October 15, 2004 Research Qualifying Exam 17
Results: NEAT-GRN Number of Signals Results averaged over 200 runs; Same GRN parameters October 15, 2004 Research Qualifying Exam 18
Results: NEAT-GRN Number of Signals Results averaged over 200 runs; Same GRN parameters October 15, 2004 Research Qualifying Exam 19
Results: NEAT-GRN XOR & NXOR Results averaged over 100 runs 35% solution success (max 250 gen) October 15, 2004 Research Qualifying Exam 20
Results: XOR & NXOR network bias_t inpt_t outp_t hidn_t hidn_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t [000] [001] [002] [003] [004] [005] [007] [011] [013] [019] [000] [001] [002] [003] [004] [005] [006] [007] [008] [009] [010] [013] [014] [017] [018] [021] [022] [025] [026] [029] [030] [038] [042] [043] [044] [046] [053] [054] [055] [060] [064] [078] [079] [080] [082] [083] [085] [093] [094] [095] 0. 000 1. 000 0. 500 r{ + 1/0. 029 0. 500 r{ - 0/0. 022 0. 250 r{ - 0/0. 030 0. 125 r{ + 2/0. 046 0. 625 r{ - 1/0. 012 0. 313 r{ + 3/0. 013 [e] 0 3 5. 88 r{ [e] 2 3 -3. 90 r{ [e] 2 3 8. 10 r{ [e] 1 3 -6. 04 r{ [e] 2 4 2. 70 r{ [e] 4 3 11. 65 r{ [e] 1 5 -9. 92 r{ [e] 1 5 -10. 55 r{ [e] 1 5 -6. 01 r{ [e] 5 3 -8. 85 r{ [e] 1 4 6. 69 r{ [e] 0 4 -1. 19 r{ [e] 0 4 8. 59 r{ [e] 0 4 -4. 49 r{ [e] 0 5 1. 68 r{ [e] 2 5 3. 76 r{ [e] 2 5 3. 41 r{ [e] 2 5 -10. 39 r{ [e] 1 7 1. 17 r{ [e] 7 5 -12. 00 r{ [e] 2 7 7. 27 r{ [e] 7 4 -1. 33 r{ [e] 7 3 5. 15 r{ [e] 7 3 2. 48 r{ [e] 0 7 -3. 24 r{ [e] 0 7 7. 14 r{ [e] 1 11 7. 35 r{ [e] 11 7 5. 03 r{ [e] 7 13 1. 27 r{ [e] 13 3 -5. 13 r{ [e] 11 3 6. 86 r{ [e] 0 11 2. 43 r{ [e] 11 5 -1. 64 r{ [e] 2 11 -0. 37 r{ [e] 11 4 -8. 51 r{ [e] 11 19 -1. 32 r{ [e] 19 5 -7. 22 r{ [e] 19 4 -8. 26 r{ [e] 19 3 -4. 63 r{ [e] 1 19 -0. 33 r{ [e] 4 13 9. 26 r{ [e] 0 13 -2. 86 r{ [e] 2 13 2. 33 r{ [e] 2 13 -3. 19 r{ [e] 11 13 7. 32 r{ [e] 5 13 -2. 39 r{ [e] 1 13 3. 85 r{ [e] 2 19 -2. 90 r{ [e] 7 19 4. 30 r{ [e] 19 13 2. 90 r{ 0. 054 } 0. 043 } 0. 049 } 0. 017 } 0. 096 } 0. 048 } - 2/0. 028 - 2/0. 007 + 0/0. 021 + 3/0. 007 - 2/0. 026 - 0/0. 035 + 1/0. 007 + 2/0. 102 + 1/0. 021 - 3/0. 081 - 1/0. 049 - 2/0. 088 - 2/0. 045 - 0/0. 012 - 3/0. 051 - 0/0. 050 + 0/0. 071 - 3/0. 054 + 3/0. 097 - 3/0. 004 + 0/0. 046 - 2/0. 023 - 2/0. 006 + 0/0. 031 - 1/0. 026 + 3/0. 050 + 3/0. 028 + 2/0. 021 - 1/0. 040 - 0/0. 020 - 2/0. 032 - 1/0. 011 + 2/0. 027 + 3/0. 013 + 0/0. 024 - 3/0. 058 - 3/0. 022 + 3/0. 076 - 1/0. 155 + 2/0. 048 + 3/0. 078 + 3/0. 021 + 1/0. 039 - 1/0. 086 + 3/0. 034 - 2/0. 029 - 0/0. 039 - 1/0. 020 + 1/0. 001 - 2/0. 046 Env 0 0. 039 0. 037 0. 005 0. 008 0. 036 0. 065 0. 040 0. 024 0. 010 0. 050 0. 005 0. 019 0. 034 0. 051 0. 008 0. 026 0. 067 0. 022 0. 045 0. 040 0. 059 0. 030 0. 064 0. 070 0. 021 0. 030 0. 050 0. 041 0. 043 0. 027 0. 003 0. 023 0. 036 0. 074 0. 105 0. 031 0. 021 0. 016 0. 029 0. 009 0. 085 0. 061 0. 042 0. 006 0. 038 0. 021 0. 056 0. 025 } } } } } } } } } } } } } October 15, 2004 Env 1 Research Qualifying Exam 21
Outline • Motivation and Related Work • System Overview and Results • Conclusions October 15, 2004 Research Qualifying Exam 22
Conclusions: Contributions • A GRN model that features a variablydecoding phenotype • 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 October 15, 2004 Research Qualifying Exam 23
Conclusions: Cars bias_t inpt_t outp_t hidn_t hidn_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t link_t [000] [001] [002] [003] [004] [005] [007] [011] [013] [019] [000] [001] [002] [003] [004] [005] [006] [007] [008] [009] [010] [013] [014] [017] [018] [021] [022] [025] [026] [029] [030] [038] [042] [043] [044] [046] [053] [054] [055] [060] [064] [078] [079] [080] [082] [083] [085] [093] [094] [095] 0. 000 1. 000 0. 500 r{ + 1/0. 029 0. 500 r{ - 0/0. 022 0. 250 r{ - 0/0. 030 0. 125 r{ + 2/0. 046 0. 625 r{ - 1/0. 012 0. 313 r{ + 3/0. 013 [e] 0 3 5. 88 r{ [e] 2 3 -3. 90 r{ [e] 2 3 8. 10 r{ [e] 1 3 -6. 04 r{ [e] 2 4 2. 70 r{ [e] 4 3 11. 65 r{ [e] 1 5 -9. 92 r{ [e] 1 5 -10. 55 r{ [e] 1 5 -6. 01 r{ [e] 5 3 -8. 85 r{ [e] 1 4 6. 69 r{ [e] 0 4 -1. 19 r{ [e] 0 4 8. 59 r{ [e] 0 4 -4. 49 r{ [e] 0 5 1. 68 r{ [e] 2 5 3. 76 r{ [e] 2 5 3. 41 r{ [e] 2 5 -10. 39 r{ [e] 1 7 1. 17 r{ [e] 7 5 -12. 00 r{ [e] 2 7 7. 27 r{ [e] 7 4 -1. 33 r{ [e] 7 3 5. 15 r{ [e] 7 3 2. 48 r{ [e] 0 7 -3. 24 r{ [e] 0 7 7. 14 r{ [e] 1 11 7. 35 r{ [e] 11 7 5. 03 r{ [e] 7 13 1. 27 r{ [e] 13 3 -5. 13 r{ [e] 11 3 6. 86 r{ [e] 0 11 2. 43 r{ [e] 11 5 -1. 64 r{ [e] 2 11 -0. 37 r{ [e] 11 4 -8. 51 r{ [e] 11 19 -1. 32 r{ [e] 19 5 -7. 22 r{ [e] 19 4 -8. 26 r{ [e] 19 3 -4. 63 r{ [e] 1 19 -0. 33 r{ [e] 4 13 9. 26 r{ [e] 0 13 -2. 86 r{ [e] 2 13 2. 33 r{ [e] 2 13 -3. 19 r{ [e] 11 13 7. 32 r{ [e] 5 13 -2. 39 r{ [e] 1 13 3. 85 r{ [e] 2 19 -2. 90 r{ [e] 7 19 4. 30 r{ [e] 19 13 2. 90 r{ 0. 054 } 0. 043 } 0. 049 } 0. 017 } 0. 096 } 0. 048 } - 2/0. 028 - 2/0. 007 + 0/0. 021 + 3/0. 007 - 2/0. 026 - 0/0. 035 + 1/0. 007 + 2/0. 102 + 1/0. 021 - 3/0. 081 - 1/0. 049 - 2/0. 088 - 2/0. 045 - 0/0. 012 - 3/0. 051 - 0/0. 050 + 0/0. 071 - 3/0. 054 + 3/0. 097 - 3/0. 004 + 0/0. 046 - 2/0. 023 - 2/0. 006 + 0/0. 031 - 1/0. 026 + 3/0. 050 + 3/0. 028 + 2/0. 021 - 1/0. 040 - 0/0. 020 - 2/0. 032 - 1/0. 011 + 2/0. 027 + 3/0. 013 + 0/0. 024 - 3/0. 058 - 3/0. 022 + 3/0. 076 - 1/0. 155 + 2/0. 048 + 3/0. 078 + 3/0. 021 + 1/0. 039 - 1/0. 086 + 3/0. 034 - 2/0. 029 - 0/0. 039 - 1/0. 020 + 1/0. 001 - 2/0. 046 Env 0 0. 039 0. 037 0. 005 0. 008 0. 036 0. 065 0. 040 0. 024 0. 010 0. 050 0. 005 0. 019 0. 034 0. 051 0. 008 0. 026 0. 067 0. 022 0. 045 0. 040 0. 059 0. 030 0. 064 0. 070 0. 021 0. 030 0. 050 0. 041 0. 043 0. 027 0. 003 0. 023 0. 036 0. 074 0. 105 0. 031 0. 021 0. 016 0. 029 0. 009 0. 085 0. 061 0. 042 0. 006 0. 038 0. 021 0. 056 0. 025 } } } } } } } } } } } } } October 15, 2004 Env 1 Research Qualifying Exam 24
Conclusions: Next Robots October 15, 2004 Research Qualifying Exam 25
Long-term Objectives: Project Overview An outline of the work to be done between now and October ‘ 05 I. Academic a. Literature search / reading b. Qualifying examination c. Thesis proposal d. Doctoral dissertation II. Robotic platform a. Design and fabrication b. Robot chassis and motor system c. Sensors and cameras d. Firmware and drivers October 15, 2004 III. Software a. Artificial brain modules: i. NEATer with GRN ii. NEATer with development iii. NEATer with topology iv. Synthetic Brains (integrated) b. Simulation and evolution: i. Simulated arm and motors ii. Simulated sensors iii. Evolutionary algorithm Research Qualifying Exam 26
- Slides: 26