Synthetic Brains: Update

Transcription

1 Synthetic Brains: Update Bryan Adams Computer Science and Artificial Intelligence Laboratory (CSAIL) Massachusetts Institute of Technology Project Review January 04 through April 04 Project Status Current research: May 04 to present Short-term objectives Project Future Long-term objectives Talk Outline June 25, 2004 Toyota-MIT Artificial Brains Project 2 1

2 Project Review: Objective Goal: A biologically-inspired robot controller Create robust behavior on a simple robotic platform Discover underlying principles that can be applied to conventional controllers June 25, 2004 Toyota-MIT Artificial Brains Project 3 Project Review: Approach Engineering + Functional + Elegant + Biology + Robust + Adaptive Brittle Limited Synthetic Brains Complex Inefficient June 25, 2004 Toyota-MIT Artificial Brains Project 4 2

3 Project Review: System From January 04: Synthetic Brains Included three biologically-based innovations 1. Genetic Regulatory Network (GRN) 2. Cellular Development 3. Complex Neural Topology June 25, 2004 Toyota-MIT Artificial Brains Project 5 Project Review: April 04 Results Each individual component performed as expected but brains failed to evolve into functional controllers. Impossible to test three experimental systems simultaneously Too much biology, not enough engineering June 25, 2004 Toyota-MIT Artificial Brains Project 6 3

4 Project Review: Accomplishments Other milestones Arm testing platform Simulation Robot GUIs Visualize protein levels Complex neural topologies Thesis proposal June 25, 2004 Developmental Neural Networks for Robust Robotic Behavior Toyota-MIT Artificial Brains Project 7 Talk Outline Project Review January 04 through April 04 Project Status Current research: May 04 to present Short-term objectives Project Future Long-term objectives June 25, 2004 Toyota-MIT Artificial Brains Project 8 4

5 Engineering Practicality Current Research: Task Space (NEAT) (Goal)? Synthetic Brains Biological Fidelity June 25, 2004 Toyota-MIT Artificial Brains Project 9 Engineering: Standard, three-layer neural net + Backwards connections + Improved evolutionary algorithm (speciation, innovation protection) Current Research: NEAT NEAT: NeuroEvolution of Augmented Topologies Biology Direct genetic encoding Static topology Static weights June 25, 2004 Toyota-MIT Artificial Brains Project 10 5

6 Current Research: NEAT Publications Double inverted pole-balancing task (Stanley & Miikkulainen, 2001) Final solution: 4 input nodes, 1 hidden node, 1 output node, 6 links Transfer from simulation to unstable domain (Gomez & Miikkulainen, 2003) Double pole balancing in simulation, evaluated on mathematical model of an unstable system A roving eye for GO (Stanley & Miikkulainen, GECCO 2004) Able to beat Gnugo on 5x5, less effective on 7x7 Competing simulated robot controllers (Stanley & Miikkulainen, 2004) Two simulated Khepera robots compete for food June 25, 2004 Toyota-MIT Artificial Brains Project 11 Current Research: Implementation NEAT Experiment NEAT implementation Robot arm environment Input from camera Output to arm servos Evolved in simulation Used to control robot June 25, 2004 Toyota-MIT Artificial Brains Project 12 6

7 Current Research: NEAT task Track an object in the visual field 1 4 2DOF (pan, tilt) Simulated CMU cam, pointing forward, on the end of the arm 80x143 resolution 15ºhoriz, 20ºvert 2 3 Evaluation: target starts in center, moves diagonally to one quadrant corner, repeat for each corner June 25, 2004 Toyota-MIT Artificial Brains Project 13 Current Research: Fitness Fitness function: ƒ = α {Φ(x pos ) + Φ(y pos ) } β (x dist ) where (x pos, y pos ) is the target s position in the camera s FOV Good Bad Φ(p) = -abs((p-p max )/p max ) + 1 p max = max camera resolution horiz = 80 vert = Φ(p) p max/2 p max June 25, 2004 Toyota-MIT Artificial Brains Project 14 7

8 Current Research: Results NEAT Experiment Results Max Fitness Generations Run 1 (Seed 1234) Run 2 (Seed 2199) Run 3 (Seed 4498) Run 4 (Seed 10000) June 25, 2004 Toyota-MIT Artificial Brains Project 15 Current Research: Results NEAT Experiment Results Mean Fitness Generations Run 1 (Seed 1234) Run 2 (Seed 2199) Run 3 (Seed 4498) Run 4 (Seed 10000) June 25, 2004 Toyota-MIT Artificial Brains Project 16 8

9 Current Research: Winning Network MyNEAT does not have speciation or innovation protection, so nets are not minimal Does reach maximal fitness (marginally better than hand-designed) June 25, 2004 Toyota-MIT Artificial Brains Project 17 Current Research: On the Robot [robot_tracking.avi] June 25, 2004 Toyota-MIT Artificial Brains Project 18 9

10 Project Review January 04 through April 04 Project Status Current research: May 04 to present Short-term objectives Project Future Long-term objectives Talk Outline June 25, 2004 Toyota-MIT Artificial Brains Project 19 Short-term Objectives: NEAT In order to test each part of the experimental system, integrate it with a known system NEAT (Stanley & Miikkulainen, 2001) June 25, 2004 Toyota-MIT Artificial Brains Project 20 10

11 Short-term Objectives: NEATer NEAT for evolutionary robotics (NEATer) NEAT NEATer with GRN NEATer with development NEATer with cell topology June 25, 2004 Toyota-MIT Artificial Brains Project 21 Short-term Objectives: Why NEATer? Although CE (cellular encoding) demonstrates that it is possible to evolve developmental systems, we chose direct encoding for NEAT because, as Braun and Weisbrod (1993) argue, indirect encoding requires more detailed knowledge of genetic and neural mechanisms. In other words, because indirect encodings do not map directly to their phenotypes, they implicitly restrict the search to the class of topologies to which they can be expanded Stanley & Miikkulainen, June 2001 June 25, 2004 Toyota-MIT Artificial Brains Project 22 11

12 Short-term Objectives: Why NEATer? NEATer is an attempt to use more detailed knowledge of genetic and neural mechanisms to expand the class of nets and create robust robotic behavior. June 25, 2004 Toyota-MIT Artificial Brains Project 23 Short-term Objectives: Task Space Engineering Practicality (NEAT) (Goal) NEATer Synthetic Brains Biological Fidelity June 25, 2004 Toyota-MIT Artificial Brains Project 24 12

13 Short-term Objectives: Goals NEAT Efficiency Speed Minimal nets NEATer Robustness Complex behavior Scalability June 25, 2004 Toyota-MIT Artificial Brains Project 25 Short-term Objectives: GRN GRN is decoded into a description of a NEAT network Promoters, Enhancers, Suppressors Protein concentrations create network configuration ugaaugcgcguuaagaga promoter val lys Task: Evolve a controller that is robust to bad lighting, disabled motors, etc. June 25, 2004 Toyota-MIT Artificial Brains Project 26 13

14 Short-term Objectives: Development Developmental step allows prefigured network to change Cells communicate with local (chemical) signaling mechanisms Signals can change weights, topology Task: Evolve a developing controller, turn camera upside down, regain behavior without further evolution June 25, 2004 Toyota-MIT Artificial Brains Project 27 Short-term Objectives: Topology Remove three-layer restriction, give cells large grid Any cell can connect to any other cell Locality: cells have a fixed location, set of neighbors Task: Perform more complex behavior, like tracking two (different) targets of different color June 25, 2004 Toyota-MIT Artificial Brains Project 28 14

15 Project Review January 04 through April 04 Project Status Current research: May 04 to present Short-term objectives Project Future Long-term objectives Talk Outline June 25, 2004 Toyota-MIT Artificial Brains Project 29 Long-term Objectives: NEATer Once the three NEATer experiments have been conducted, the final step will be to integrate them into a single system June 25, 2004 Toyota-MIT Artificial Brains Project 30 15

16 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 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 June 25, 2004 Toyota-MIT Artificial Brains Project 31 16