Al OVERVIEW Samuel M. Smith Ph.D. 2018.02.08 sam.smith@consensys.net
Computing techniques that attempt to replicate various aspects of human perception and cognition that is only taught to graduate students.
Weak or Narrow A.l.
Strong or General A.l.
Many Types
Automated reasoning Reinforcement Learning Machine Learning Meta learning Machine Intelligence Deep Learning Computational Intelligence Artificial Neural Networks
Fuzzy Logic Artificial General Intelligence v LoB
Genetic Algorithms Natural Language Processing
Evolutionary Algorithms Computational Search Bayesian Belief Networks Pattern Recognition Symbolic Reasoning Image Processing Approximate Dynamic Programming Intelligent Control Expert Systems
Machine Intelligence Landscape
MACHINE INTELLIGENCE 3.0
EE ew mEE TED ZE Eeoas
ImmE ST | moEE W MM it
o — A S S S
AT | Dhomn | SRR || ommpm | Do @rmems || Tamor e renbd
= = =
BRI e | Eme el | TED | olmen tee shivonzilis.com/MACHINEINTELLIGENCE - Bloamberg BETA = . G
The Many Tribes of Al Machine B The Many Tribes of (& ‘1 e B O | | j | L EE’” _J) ol
Deep Learning Supervised learning using multi-layer back-propagation, stochastic gradient descent, on pre-classified training data. Most popular. Many open source tool kits. Works best with very large data sets. Simple algorithm Relatively easy to implement and use Primary application is pattern recognition or image classification
Hierarchically Decomposed Staged Concurrent Modular Fuzzy Learning Algorithm Defeated Human Every Time in Simulated Air-to-Air Combat Genetic Fuzzy based Artificial Intelligence for Unmanned Combat Aerial Vehicle Control in Simulated Air Combat Missions. 2016 Emst et al Journal of Defense Management Volume 6 » Issue 1 http://dx doi.org/10.4172/2167-0374.1000144
Basic Machine Learning Types Supervised Learning - Classified Data Set Unsupervised Learning - Clustering of Unclassified Data Set Reinforcement Learning - Adaptive Critic = | IR ~ s ) AN : _ 3 AT E TN = | || ey 2 e 1 SO\
Basic Machine Learning Types Supervised Learning - Classified Data Set Unsupervised Learning - Clustering of Unclassified Data Set Reinforcement Learning - Adaptive Critic }“:rfi:" .‘H = S, - ‘ZZZ, 3 , ’3;:. ”:;."'4:.’5_:‘:, N
Basic Machine Learning Types
Supervised Learning - Classified Data Set Unsupervised Learning - Clustering of Unclassified Data Set Reinforcement Learning - Adaptive Critic
Reward R
State i
Value Action [ J
Common Concepts Stochastic Gradient Descent: Continuous geometric function from input to output with continuous derivative Universal Function Approximators: Neural net, fuzzy inference system, etc No Free Lunch Theorem All algorithms that search for an extremum of a cost function perform exactly the same when averaged over all possible cost functions. So, for any search/optimization algorithm, any elevated performance over one class of problems is exactly paid for in performance over another class. Proficiency of practitioner is often more important than the method used.
Stochastic Gradient Descent (w) 'W"'E'I‘;(\ ,,' —— Gradient re // i ittt i .m/ ’ w) w
New Kinds of Al for Blockchain Artificial Social Intelligence Artificial Network Intelligence Group Al Decentralized Al Distributed Al All need identity, reputation and distributed consensus
Distributed Consensus Proof of Work Proof of Stake Byzantine Agreement Coordinated Control
Background
Autonomous Underwater Vehicles L i o gt~ B
Environmental Adaptive Sampling anc=="- ) VAR - o i . f N\ \ o cmen e o S \ E— L\ \ e ' \ \ | I B s N P \ 5 K e . e N\
Shipboard & Building Automation I A | 247N AL e ST e, Al B = e = S filkews. ¥ By
Reinforcement Learning: = Heuristic Dynamic Programming Computational Search
e Cell-State- Space-Based ) [ Search
! — .
B et et I IR i (s ) SEEmpT SR an B
PHYSICAL SECURITY OF CRITICAL INFRASTRUCTURE pre [ ! Y 3 M P i L - 4 b V. {
Strategic, Tactical, Executive Hybrid Layered Control
Deliberative, Compositive, Reactive Coni [ Cogans
Sy Sy Svsems
Hybrid Hietarchial Control a0 )
Dperete [Tk Stnategic Execution Tactical Strategic Strategic Tactical Execution
sy fi g CoptlinPlty PaisCordien S0 AN WA Control Tactical Laper,
Continuum {Coordimtion i
e
e M L e e e Simple
Coordinated Control - Distributed Consensus Groups of autonomous agents from different circumstances cooperating to achieve shared objective(s) in an environment of uncertainty Think Decentralized Autonomous Organizations Think Distributed Autonomic Services
Coordinated Control Types Leader-Follower Behavioral Role Based (Virtual Structure)
Leader Follower Full state of the leader is broadcast to followers Leader sequences group tasks Followers execute a “formation” relative to the leader Allows mathematical analysis and performance guarantees Centralized Fragility to communication failure and member loss Heterogenous behavior difficult
Behavioral Each agent follows prescribed behavior rules relative to environment. Several “behaviors” are prescribed for each “agent” or vehicle and the Control is derived from a weighted combination of the behaviors Difficult to analyze mathematically or guarantee performance Decentralized Robust to communication failure and member loss Heavily dependent on environmental coherence Precise coordination difficult Swarming, flocking