วีดีโอ

Very nice lecture, thank you!

Thank you.

Nice explanation you can also say that # - acts as an placeholder

1. Introduction to Genetic Programming Overview of genetic programming as an evolutionary algorithm for software development. Mention of genetic algorithms, evolutionary strategies, and evolutionary programming discussed previously. Reference to genetic programming in the context of artificial life. 2. Representation of Solutions Discussion on the use of a tree structure to represent solutions in genetic programming. Introduction to the population consisting of expressions with functions and terminals. Adaptation of sets of functions and terminals based on the problem being solved. 3. Challenges in Genetic Programming Challenges of closure (safe function execution) and sufficiency (providing adequate building blocks) in genetic programming. Consideration of different representations of programs, such as binary, pseudo code, programming language code, or diagrams. 4. Random Expression Generation Methods for generating random expressions, including full, grow, and ramped half-and-half approaches. Explanation of how these methods contribute to diversity in the initial population. 5. Mutation and Crossover Operations Description of mutation and crossover operations in genetic programming. Use of cutting and adding a new random subtree for mutation and exchanging subtrees between parents for crossover. 6. Handling Uncontrolled Growth Measures to avoid uncontrolled bloating of expressions, such as introducing penalties for size or depth limits. Importance of defining safe operations to interpret or evaluate generated programs in a virtual environment. 7. Symbolic Regression Introduction to symbolic regression as a notable application of genetic programming. Definition of symbolic regression as finding expressions to approximate sets of points. 8. Fitness Landscape Challenges Discussion on the rugged and non-continuous fitness landscape in genetic programming. Consideration of symbolic and discretized nature of solutions contributing to the challenge. 9. Efficiency Concerns in Optimization Questioning the efficiency of genetic programming in solving optimization problems due to the complex fitness landscape. Mentioning ongoing efforts in the scientific world to improve the efficiency of the genetic programming approach. PART 2: - Symbolic Regression in Genetic Programming (GP): - Symbolic regression is a key application of Genetic Programming. - Unlike traditional regression methods where the form of the function is fixed and only coefficients are sought, GP allows for manipulation of the function's form itself. - In symbolic regression, the goal is to find a function that accurately describes a given set of points. - Form of the Expression in Symbolic Regression: - The form of the expression sought is controlled by selecting elements in the set of functions (denoted as "f" and "t") and imposing potential restrictions. - The text mentions two links to educational programs with example Python code illustrating the concept. - Example Experiment 1 - Symbolic Regression for a Test Function: - The example involves finding an expression that best describes a set of points corresponding to an unknown function. - GP is given access to variables (X), operators (negation), and constants (-1, 0, 1). - Through evolution, GP refines its population, and the fitness of each individual (expression) is assessed based on how well it approximates the expected values. - Example Experiment 2 - Evolution of Logic Circuits for XOR: - The experiment aims to evolve a logic circuit that implements the XOR function. - Different sets of functions (negated, and, if-then) are explored, and GP evolves solutions by combining these functions. - The text discusses variations in success based on the functions allowed during evolution. - siscussion Points: - **Expression Simplification During Evolution: - Raises the question of whether it would be beneficial to simplify expressions during evolution. - Discusses examples where expressions include redundant operations (e.g., multiplying by zero) and questions the benefits of simplification. -Power and Challenges of Genetic Programming: - GP can be applied in various domains, especially where humans lack experience. - The flexibility of GP allows for the evolution of solutions for different problems. - Hyperheuristics and Self-Programming Algorithms: - Introduces the idea of evolving optimization algorithms themselves using GP. - The structure of evolutionary algorithms, including selection techniques and crossover, can be controlled by GP. - Discusses the computational cost of evaluating and evolving such algorithms. - Conclusion and Upcoming Topics: - Concludes the discussion on Genetic Programming. - Teases the next topic on classifier systems, describing it as a mix of a simple cognitive architecture and evolutionary programming.

can you please upload a video showing EA in real time applications like earlier you used a simulator to show things like Simulated annealing and Tabu search, it would be really good to see EA that how it processes data in real time

Key Points: 1. Evolutionary Programming Concept: Opposes the idea of encoding everything in binary, emphasizing the use of natural representations. 2. Population Structure Change: From matrices of zeros and ones to vectors representing real solutions (numbers, neural networks). 3. Advantages of Natural Representation: - Better performance, as there's no need to encode/decode from binary. - More efficient optimization landscape control. 4. Differences from Genetic Algorithms: - No constraint on representation. - Genetic algorithms often involve encoding solutions as strings of tokens. 5. Fitness Landscape Smoothness: - Importance of smooth fitness landscapes for optimization efficiency. - Introduction of Fitness Distance Correlation (FDC) as a measure of smoothness. 6.Embryogeny: - Development from genotype to phenotype. - Distinction between space of solutions and space of actual solutions. 7. Mapping Genotypes to Phenotypes: - Importance in optimization of complex solutions like robots, cars, etc. - Different mappings result in different fitness landscapes. 8. Reasons to Use Non-Trivial Mapping: - Reduction of search space. - Better enumeration of the search space. - Improved constraint handling. - Compression, repetition, adaptation. 9. Challenges in Evolving Embryogeny: - Bloat (solution becoming too large). - Epistasis (complex gene interactions). - Excessive disruption of child solutions. 10. Importance of Inheritance: - Ensuring that child solutions inherit sufficient information from parents. 11. Conclusion: - Understanding the complexity of representations and their impact on optimization. - Link between the space of genotypes, phenotypes, and fitness landscapes.

Summary and key points for those of you who feel they haven't covered enough:) 1. Introduction to Evolutionary Strategies: - Evolutionary strategies are a type of evolutionary algorithm used for continuous numerical optimization. - Unlike genetic algorithms, evolutionary strategies represent solutions as vectors of continuous numbers. 2. Representation of Solutions: - Solutions in evolutionary strategies consist of vectors of numbers, representing continuous variables in mathematical optimization. 3. Genetic Operators for Continuous Space: - Discussion on how to define mutation and crossing over for solutions represented as vectors of numbers. - Focus on proposing genetic operators suitable for continuous spaces. 4. Crossing Over in Evolutionary Strategies: - Linear combination as a general approach for crossing over vectors of numbers. - Introduction of the concept of weights in linear combination, allowing for variability in the crossing-over process. 5. Mutation in Evolutionary Strategies: - Mutation involves adding a normally distributed, mean-zero, random value to each gene or element in the vector. - The concept of "creep mutation" is introduced, emphasizing slow changes to each value. 6. Normal Distribution for Mutation: - Explanation of how the normal distribution is utilized for generating random values in mutation. - Emphasis on the bell-shaped probability distribution, with smaller changes being more common and larger changes occurring more rarely. 7. Employing Evolutionary Strategies for Numerical Optimization: - The essentials for employing evolutionary strategies in numerical optimization: - Evaluation function - Continuous multi-dimensional solution space - Mutation - Crossing over - Selection methods 8. No Explicit Resolution: - Unlike previous examples with low resolution, evolutionary strategies use floating-point numbers without specifying a specific resolution. 9. Conclusion: - Evolutionary strategies provide a mathematical and abstract approach to solving optimization problems in continuous spaces. - Floating-point numbers are used without low-resolution assumptions, making it applicable to various numerical optimization scenarios.

Here is a quick summary for those of you who didnt actually understand the main idea of this video😃: Introduction to Biologically Inspired Algorithms: Discussion on biologically inspired algorithms in artificial life, artificial intelligence, and computer science. Previous discussions covered selection in evolutionary algorithms and the main loop of the algorithm. Crossing Over and Mutation: Introduction to the crossing over operator in evolutionary algorithms. Questioning whether crossing over is mandatory for the algorithm to work efficiently. Mention of the analogous operation in other algorithms, termed as "mutation." Necessity of Crossing Over: The discussion on whether crossing over is necessary for the algorithm to function efficiently. Emphasis on the importance of assessing whether crossing over adds value and whether it fulfills the requirements for an efficient operation. Mention of various types of crossing over operations, such as single point crossover and uniform crossover. Challenges of Crossing Over: Highlighting that crossing over creates new solutions by inheriting information from two or more parent solutions. Challenges in implementation: merging information from diverse parents to create a valid and effective child solution. Mention that crossing over may sometimes work like a large mutation, potentially teleporting solutions to unrelated regions. Mutation Operator: Introduction to the mutation operator, which is likened to the neighborhood operator. Questioning the necessity of the mutation operator when crossing over is present. Emphasis on the role of mutation in introducing diversity to the population and preventing premature convergence. Importance of Mutation: Mutation is considered essential unless crossing over fulfills the role of introducing sufficient diversity. Mutation is crucial for traversing the fitness landscape and discovering optimal solutions. Comparison of Crossing Over and Mutation: Crossing over is useful when a child solution should be similar to parents, fulfilling specific requirements. Mutation creates a new solution inheriting information from a single parent, preventing convergence and introducing variability. Conclusion: The importance of both crossing over and mutation in evolutionary algorithms. Acknowledgment that the discussion on crossing over and mutation will continue when exploring representations other than binary, such as vectors of real numbers.

To find the other robots in the football field, I would say geographically sourced reference information, e.g. knowing where your goal is and from which direction you came from before, would allow you to know which side you are on. Other reference points such as colors in the jerseys of the robots or a slight detail in the goal might help you discern between different goals. In general, finding as many reference points to differentiate both sides of the game will enable you to see them as distinct and take better decisions.

Interesting! any plans for updates on this series of videos? the research looks pretty interesting!

Thank you very much for this lucid explanation of classifier systems.

Thanks for the video. It helped me get some concepts better. Keep the good work Maciej, greetings.

Thank you very much teacher

Good stuff! There's a few more now, maybe an update video? 🤔

Hey! Have you tried genetic+Reinforcement learning for trading ?

There is a martial art taught to Russia's military elite called 'Systema'(aka 'the system'). Its like the perfected version of everything you describe about cognitive architectures. It doesn't teach any techniques. It teaches a system of thinking that focuses on the rate in which someone can create, reset, delete, and adapt their cellular memory mappings to move perfectly regardless of situation or scenario. Quite a few studies were done on master practitioners of 'systema'. Giuseppe Filotto(himself a master in it) wrote a book about it called "Systema The Russian Martial System: Created by the Soviet Military for Their Special Forces Elite" that I suspect you'd find fascinating in its similarities to some of the cognitive architectures of AI, but overall, imo, its a cognitive architecture(system of thinking and acting) thats far superior to every cognitive architecture created or postulated for AI. It likely has remained hidden from the academic world because its a unique martial art taught to elite military units(and was kept secret for years because of old Soviet government).

Loved the explanation, you made alot of things click into position for me. But I disagree with your opinion on aspirations for global optimum not being natural. The 100 monkey experiment explains this type of communication well. As does the world wide WEB(web of communication that spans the entire global population). Imo pheromones is such a poor term. I wish science didn't use the term pheromones, its a 'quantitative' term for a 'qualitative' topic. Its like when 'star wars' used the term 'midichlorians'. It just doesn't fit. Its more like a gut feeling or intuition (or 'the force'). Its more of a 'qualitative' thing than 'quantitative'. For most animals and insects their 'gut feeling' or intuition is farrrr more in tune with nature than humans. Whenever there is a natural disaster like a tsunami, animals and insects sense it hours before humans and rarely do any animals or insects actually perish. That's not the effect of pheromones. Its like a sixth sense(really its our first sense), intuition, the force, the universal being of the universe, the one the only. We are both all for one(local as in local to each and every organ within our bodies) and one for all(global as in global to each and every organism within the globe/earth), we are both and it is infinitely large and small. The line from the matrix "There is no spoon" is an accurate statement. PSO is the model for real life. We dont have a soul, we ARE a soul inhabiting an inertial physical body in a L-system called Life, L stands for Life not Lindenmeyer otherwise they'd have called it the Lindemeyer system. Inertia is the weight of the physical body we are currently inhabiting(weight is dependent on 7 different factors). Aspiration for local optimum is equivalent to service to self, exploitation, all for one, or human organs make up the human. Aspiration for global optimum is equivalent to service to others, exploration, one for all, or earth organisms make up the earth.

Well and easy explained! Thank you.

Thank you Maciej Komosinski for this awesome explaining of Evolutionary Computing for Global optimization, it was very useful and inspiring. no many visual resources for this topic, appreciate your contribution. you delivered it in a simple manner, and OptVis tool is also awesome. also the Polish accent is wonderful (:

I couldn't reach the script can u plz share by another way thanks

Thank you so much Prof. Komosinski, these are amazing videos!, just what I have been looking for.

In what route optimization scenario or application, would you actually want a "high variable" for inertia if your goal is for faster route convergence for a "global consensus"?

Extremely well defined hierarchical relations. Regarding AI safety at 08:58: autonomous robots will need some sort of value system working as some sort of feedback loops. If this system is "read-only" and programmed by humans then intellectual robots will be safe. The ethical values may be stored in the highest level of hierarchy so they can block any other layer, I think. Some highly ethical people have this kind of control, which does not deprive them from freedom and does not make them less intelligent (there is quite a massive sub-culture in humans though who believe that doing good is very"stupid")

If Tabu Search is always willing to find new wouldn't it act like Greedy more than Steepest?

If we use the different fitness functions, increasing the rate of mutations to maintain a diverse population of solution what is the probability that if we are scarfing short term fitness to obtain longer term fitness. Is there any chance that it can't convert into longer term fitness as the "no free lunch theorem proves that there is no general solution to this problem.

Question: what if I am interested in finding the local minima on that 3d plot (multi dimensional landscape). How can I find the local minima?

Hi, are we supposed to send our homework or is it just for our personal development?

Since we consider individual cells to be "life", are animals and plants "life" or just a system of cells?

Life - śmiertelna chorob przenoszona drogą płciową

First comment !!(hahahah)

Nice, thanks a lot ;)

I am very appreciate for the tutorial you posted!!!