CHALLENGES IN DESIGNING ROBOTIC BRAINS

Transcription

1 CHALLENGES IN DESIGNING ROBOTIC BRAINS Dr. Jeff Buechner Department of Philosophy Rutgers University-Newark Director Rutgers-Merck Summer Bioethics Institute 2012

2 Robotics is Interdisciplinary Designing robots requires knowledge from physics, chemistry, biology, psychology especially cognitive science, computer science, neurosciences, engineering, mathematics, logic, philosophy, anthropology, sociology, politics, education, and, not least, common sense. Common sense is perhaps the hardest part of designing robots.

3 Robotics is interdisciplinary Why do you need to know something about philosophy, anthropology, political science, sociology, education, and common-sense to design robotic brains? Philosophy: (i) Ethics do you want a robot to have the ability to engage in ethical reasoning? (ii) Is consciousness a physical phenomenon? If it is not, then there are no physical parts that you could use to design a robotic brain which is conscious.

4 Biology and physics: bio-inspired robotics All insects secrete a sticky fluid from their feet as they walk, which creates a liquid bridge between their feet and the walking surface, forming a glue-like strong bond through surface and molecular tension. Insect-like robots (insect-bots) are climbing robots. Grooves on their feet enhance their sticking power, by creating more fluid bridges between their feet and the walking surface.

5 Biology and physics: bio-inspired robotics Geckos (a lizard that sells auto insurance) do not use fluids on their feet to climb, but rather millions of hairs on their footpads to stick to surfaces. Hairs bend on contact with the surface, and intermolecular van der Waals forces cause the two to stick together. Gecko-bots use carbon nanotubes in place of hairs. However, gecko-bots have a hard time climbing rough, wet, or salty surfaces, since water and salt interfere with van der Waals forces.

6 Human intelligence comes from human brains A human brain has neurons (That s 100, 000,000,000 neurons) In a human brain there are synapses (neuron-neuron connections) (That s 100,000,000,000,000 connections!) FACT: In 2007 all the general computers in the world could together perform 6.4 x instructions per second. That roughly equals the number of nerve impulses produced by one human brain each second.

7 WATSON beats humans on Jeopardy! A few weeks ago a robotic brain WATSON beat two human contestants on the quiz show Jeopardy! WATSON does not forget, and it has a much larger working memory space than humans. Evolution created human working memory and the rate of retention and the rate of forgetting is probably optimal for us. We can design robotic brains with different retention and forgetting rates. But we do not understand how these rates in human brains affect human intelligence. WATSON also lacks the kind of common-sense that is part of human intelligence. It is sometimes baffled by simple questions and it thinks fictional characters are real.

8 Bio-inspired robotic brains neural networks Robotic brains which are modelled on human brains are called neural networks. Neuroscientists still do not know what are the basic neural circuits of the human brain, and how they relate to one another. Work in the neurosciences goes hand-in-hand with work on neural networks. A huge project is now underway: mapping the human brain the connectome project.

9 Project: Design a robotic hand You need to think about what animal hands (such as human hands) do and how they do it that is, what functions do human hands perform and how are those functions achieved? Your answer is not independent of the environment in which a hand performs its actions

10 Hands in action How would you tell the hand how much pressure to put into gripping a (i) human being, (ii) another robot, (iii) a piece of concrete that weighs 1,000 pounds, (iv) a piece of concrete that weighs 1/1000 th of an ounce? How would you tell the hand how to move in 3-dimensional space to get to its intended target? (This involves physics and mathematics. Higher-dimensional shapes are useful for motion planning in robots. The path of a hand that moves in 3-D can be represented as a higher-dimensional shape. These are smooth shapes written as differential equations which allows us to describe them in terms of their flow. ) Does the shape of your grip change as circumstances change? Is your grip the same shape when you use a fork (to eat with) and when you hold a baseball (to throw)?

11 ROBOT BRAINS To design a robot brain you need to know what environment (what world) the brain will inhabit. Why? Because brains control how someone (something) responds to their environment. But you need more you also need to know what body the brain will inhabit. Why? Because a brain controls its body as it navigates/negotiates its environment.

12 How complex is an environment? The world in which we live is huge and complex. If a robot lives in the same world, how will its brain represent that complexity? How do human brains represent a complex and changing world?

13 A simple world BLOCKSWORLD In the early days of artificial intelligence, robotic brain designers kept the world in which the robot lived VERY simple. Suppose you are a robotic brain and your body is an arm, which can grasp blocks in one and only one way. Your environment consists of a set of three square blocks, a, b, and c, and the surface of a table on which they sit. Suppose that the movements of the arm through 3- dimensional space are controlled by the brain, and that the problem of how to get this to happen has been solved. Suppose that the actions your can perform with your arm are very limited and that your descriptions of the environment are very simple and limited.

14 BLOCKSWORLD Describing the environment: EMPTY (Arm), FULL (Arm, x), ON(x, y) and 4 objects: Blocks a, b, c, and table. If the three blocks are all on the table, and no block sits on another, then the complete description of the world is: ON(a, table) AND ON(b, table) AND ON(c, table) AND EMPTY(arm) Actions your arm can perform: (you can hold only 1 block at a time in your hand) (i) If block x has nothing on it and is not already on the table, pick it up and put it on the table (ii) If block x and block y have nothing on them, then put block x on block y. (iii) If block x is on top of block y, then take block x off of block y and put block x on the table. (iv) If block x has nothing on it, then put block x on block y.

15 BLOCKSWORLD Problem for you to solve: Block c is on top of block a, which is on the table, and block b is on the table. Arrange the blocks so that a is on top of b and b is on top of c. Before solving the problem, the complete description of the world is: 1. ON(c,a) AND ON(a, table) AND ON(b, table) AND EMPTY(arm). How do you tell your arm that you want the world to become: 2. ON(a, b) AND ON (b,c) and ON(c, table) AND EMPTY(ARM)? That is, what is your PLAN for getting the world to look like 2, starting from a world that looks like 1?

16 BLOCKSWORLD Be careful! If in your PLAN the last thing to do is to put b on c, you won t succeed, for you CANNOT do that, since a is already on b, and your arm can hold only 1 block. This means that in planning what to do, you must plan to put b on c BEFORE you put a on b. If you start planning by taking c off a, and putting a on b, you cannot succeed. To succeed you have to represent all of the possible actions you can perform, and determine which ones will not result in achieving the goal and which ones will result in achieving the goal. Even in a simple world like BLOCKSWORLD, to solve the problem, there are more than 15 possible actions to consider, some of which will depend upon what actions preceded them.

17 BLOCKSWORLD When you plan to achieve some goal, you usually break up your plan into smaller ones with smaller goals. You work on each smaller goal, and the end result is your end goal happens. Not always! Sometimes you have you know about the relations BETWEEN the work on the smaller goals. You have the smaller goal of getting block a on block b. You also have the smaller goal of getting block b on block c. But you cannot get your final goal UNLESS you WAIT to put a on b AFTER you have put b on c. You cannot work on the two smaller goals without knowing about what to look for in each. HOW DO YOU GET THIS KNOWLEDGE? HOW DO YOU KNOW THAT YOU CANNOT ACHIEVE YOUR GOAL IF YOU PUT a ON b BEFORE YOU PUT b ON c?

18 Planning your future Humans are always making plans, and our plans are certainly more complex than in BLOCKSWORLD, as is our world. How do we do it? That is, how do we figure out how to break up a goal into smaller goals, and figure out how the goals interact? The interactions can be amazingly complex. Think about cooking a meal, and all the small goals that interact with one another!

19 Sounds and Meanings A speech signal you hear has been broken into the following sounds: k a t s k a r s How do you divide those sounds into words? (i) Kat = cat skars = scares (ii) kats = cats kars = cares Which is correct? It depends on the sentence in which they occur. (i) The cat scares all the birds away. (ii) A cat s cares are few. If you did not have the additional knowledge about the structures of the whole sentences, you would not be able to group the sounds into words correctly. How do you know how much of a sentence (or paragraph, or text) to look at BEFORE you divide sounds into words?

20 Our world is complex and changing Not only is the world very complex, but it also changes. What is true at one time may become false at another time. In reasoning and planning, we have to reason about changes and be able to anticipate changes. How would you tell a robotic brain how to do this?

21 The frame problem Suppose that when you perform some action, you have a complete list of the consequences of performing that action. Is that list enough in reasoning about and planning with respect to the consequences of your actions? NO! Why not? You may know what changes, but what do you know about what does not change?

22 The frame problem You throw a ball in your living room, and it hits a lamp. The lamp falls onto the floor, and a light-bulb shatters. Pieces of glass scatter across the floor. A lot of things happened. You know what happened, because you have a list of what happened. Question: What did not happen? Did the ceiling change color? Did the refrigerator stay on? Did your foot remain attached to your body? If you are not told what does not happen (by being given a list of what does not happen), how will you ever know what does not happen when you perform an action?

23 A bad way to solve the frame problem Solution: Not only do you list the consequences of each action you perform, you also list what does not change when you perform those actions. This is bad, because a great many things do not change when you perform some action. Most of the time, more things stay the same than change. Imagine doing this while standing in the middle of a busy street, deciding whether to cross or to go back! If you had to list everything that does not change, your list would be colossal. And where do you stop? If you live in Newark, must you list every thing in Newark that does not change when you buy a soda at your local store? What about the things that change, even though you don t cause them to change? If your list says that the traffic on the street does not change when you buy a soda, is it safe to cross the street without looking? NO other changes are happening independently of the changes that occur as a result of your actions.

24 An obvious solution to the frame problem? Solution: If it is not on the list of changes, it does not change. Problems: (i) The list of changes is itself a bad way to reason about changes. It will have to be too long think of all possible changes!! Think of the things changing at this very moment! (ii) It will have to say If it is not on the list of changes, POSSIBLY it does not change. But how do you a robot brain what POSSIBLY means?

25 Jobs for good robotic brains Prediction: Given a description of the world over some period of time and a set of rules governing change (such as physics), predict the way the world will be in some future time. Explanation: Given a description of the world over some period of time, and the physics, produce a description of the world at some earlier time that, with the physics, accounts for the world being the way it is at the later time, such as now. (There is (now) a loud noise outside the building. Infer that a car crash produced it.)

26 Jobs for good robotic brains Planning: Given a description of some desired state of the world over some period of time, and of the set of rules governing change, produce a sequence of actions that would result in a world fitting that description. We need to know what must be true for the world to be a certain way, and that it is in my power to make those things true. Learning: Given a description of the world at different times, infer the set of rules governing change (the physics) which accounts for the regularities.

27 More problems! How do we reason efficiently about what is true over extended periods of time? Our physics must support inferences of the form If this is true at time t 1, then that is true at time t 2. The problem is that the if-part might get too large to be of any use to us. Suppose you have a simple physics problem: a ball rolling in a certain direction (at t 1 ) will continue rolling in that direction (at t 2 ). To infer that will happen at t 2, you need to know not just the physics, but also that there are no strong winds, that someone is not about to pick up the ball, that there are no condensed explosives attached to the ball that will explode before t 2, and so on. We could easily continue this list and make it arbitrarily large. Question: How do we reason efficiently AND reliably without having to make sure (i.e., to VERIFY) that thousands and thousands of conditions do not occur? How do we make reliable predictions about the future without taking into account EVERYTHING about the past?

28 Common sense as a solution COMMON SENSE: You do not worry that an elephant will fall on your head when you cross the street, nor that the street will open up and reveal a fiery pit below your feet. How do you tell a robotic mind what is a matter of common sense?

29 Another problem about predicting Even if we solved the previous problem, we might have to make predictions about the future very often, since the world is constantly changing, and what is true at one moment may not be true at the next moment. The shorter the time interval, the more reliable the prediction. If we predict the ball will continue rolling for an instantaneous time interval, we will have to make the prediction infinitely often between t 1 and t 2. This policy will not work we have to be less conservative about the size of the time interval. How do you tell a robotic mind how to be less conservative about the size of the time interval, so that predicting the future is more efficient and yet it is also reliable?

30 Bio-inspired intelligence Evolution gave us incredible minds to solve these problems. So why not borrow the design from nature, and use it in designing robotic minds? We can t at least we can t do it now, since we do not know how WE (human beings) do it. We do not know how we reason common-sensically. We do not know how we solve these difficult problems about reasoning, predicting,explaining, planning, and learning in a world in which changes come and go, a world that waxes and wanes over time.

31 Philosophy again! Why it might be a good idea to give a robot brain the capacity for ethical reasoning. Programming languages for robotic brains use logic for reasoning. One pitfall of logic is that anything can be inferred from a contradiction. This is easy to see. A contradiction has the form p AND not-p. (I am standing here at this location and I it is not true that I am standing here at this location. That is a contradiction.) Given p AND not-p, infer p. Then infer p OR q (by addition). Then infer not-p. Then infer q (by disjunctive syllogism).

32 A family dispute leads to tragedy! Imagine 15 years from now: you are married, have a great spouse, great job, great place to live, and a great do-it-all robot. One morning after you have been out with your friends, indulging in too many beers, you sit at the breakfast table, where there is a TV playing a crime movie. Your spouse says: You got drunk last night! You respond: I did not get drunk last night! The robot is listening to each of you. It hears a contradiction: p AND not-p. At that very moment, a gangster (in the crime movie) yells: Shoot the both of them! The robot infers that this is a valid order, and does it. If the robot had the capacity for ethical reasoning, it might not have done its dirty deed.

33 Consciousness? If a robot is conscious, we will treat it differently than if it is not conscious. (Why? Because it is more similar to a human being if conscious than if it is not conscious.) NO ONE knows whether consciousness is a physical phenomenon like water waves or sound waves or light waves or whether it is non-physical. NO SCIENTIST COULD EVER PROVE CONSCIOUSNESS IS NON-PHYSICAL, BECAUSE ALL SCIENTIFIC THEORIES ARE STATED IN TERMS OF PHYSICAL THINGS (PHYSICAL OBJECTS AND PHYSICAL PROPERTIES). If consciousness is non-physical, only a philosopher could prove that. That is the power of philosophy. But there is no philosophical proof that consciousness is non-physical, nor is there a scientific proof that it is physical. That, then, is the greatest challenge facing designers of robotic brains: to design a conscious brain.

34 More about robots! Where can you learn more about robots? The 2012 Rutgers-Merck Summer Bioethics Institute, July 8 th -July 14 th, all expenses paid. YOU ARE HERE NOW! Our focus is on ethical issues in robotics. How would you tell a robot the difference between being here now, and not being here now?