Perceptual Control Theory (PCT) as a Framework for Computer Modelling Across the Social Sciences Dr Warren Mansell Senior Lecturer School of Psychological Sciences University of Manchester Credits to Bill Powers, Tim Carey, Rick Marken, Kent McClelland, Yu Li, Savas Akgonul, Sara Tai, Martin Brown, Dominic Rogers, Eric Gruber, Christine Ihenacho, Jason Wright, Hannah Gaffney, Rachel Edwards Plan My Background What is Perceptual Control Theory (PCT)? Working examples ◦ Economics ◦ Sociology ◦ Linguistics Modelling Goal Conflict in Psychopathology MYLO – Manage Your Life Online Further discussion My Background Research into Cognitive Behavioural Therapy since 1994 Dissatisfied with cognitive & behavioural theories Came across PCT in 1998 Use for psychotherapy (Method of Levels) Basic Research – with in house computer software developer – Yu Li Aim of talk – wide applicability & demonstrate integrative capacity of PCT Demonstration Can you tell what someone is doing by watching what they are doing? ‘Rubber Band’ Demo History Self-Regulation Theory e.g. Carver & Scheier (1981) Cognitive Psychology Perceptual Control Theory William T. Powers (1960) Information theory Science Fiction e.g. ‘cyberspace’ Cybernetics Wiener (1948) Ashby (1952) Control Engineering Harold Black (1927) Homeostasis Claude Bernard (1865) Walter Cannon (1932) Ancient technology Ktesibios (c200 BC) Heron (10-70AD) Early Psychology William James (1890) John Dewey (1890s) John Dewey (1896) – the reflex arc What we have is a circuit, not an arc or broken segment of a circle. This circuit is more truly termed organic than reflex, because the motor response determines the stimulus, just as truly as sensory stimulus determines the movement. (Dewey, 1896; p. 363). History of Control Engineering Perceptual Control Theory (PCT) Developed during the 1950s by a physicist/engineer – William T. Powers First published Powers, Clark & McFarland (1960) Formalised Powers (1973) Powers’ latest Book (2008) reviewed in Nature (Mansell, 2008) Diverse range of applications published across academic domains (see www.pctweb.org) Core Principles of PCT Living organisms control their input, not their output ‘Behaviour is (merely) the control of perception’ Analogous (and possibly homologous) to homeostatic mechanisms in physiological systems Engineering principles can be used to explain these mechanisms of control Example within speech (Cziko) Make a /t/ sound Notice where your tongue is placed Repeat with tongue pressed against bottom of mouth Can you still say /t/? ability to speak comprehensibly with tongue in abnormal position shows that speech sounds are not pre-programmed motor outputs same phenomenon demonstrated when talking with cigarette, eating utensil or other object in the mouth (e.g.., food) Negative Feedback Loop Powers (2008) Basic Tracking Experiment Psychological Review (Powers, 1978) ◦ High negative correlation (-.99) between invisible disturbance (IV) and behaviour (DV) ◦ Low correlations (0.0) between disturbance and input and between input and behaviour – not a linear causal model ◦ PCT computer model provides 0.99 correlation with actual behaviour; replicated many times (Marken; Bourbon; see pctweb.org) Controlled Input (I) Human Behavior (DV) Disturbance (IV) r ~.00 IV r ~.00 I DV r ~.99 Principles of PCT Control is achieved via negative feedback Control is achieved via a specific hierarchical organisation of loops Individuals can only control their own perception; controlling others leads to conflict Conflict between high level control systems accounts for ‘dysfunction’ Reorganisation re-establishes control via a specific learning mechanism Economics: Modelling Market Agents McClelland (2010). www.pctweb.org Rational choice model of economic agents insufficient Modelled a range of risky & conservative strategies – ‘robot investors’ Relative advantages depended on economy modelled McClelland’s proposed model: Investors try to control two perceptions. Investment growth They want to see steady growth in the value of their investments. Liquidity protection They want to see steady growth (or no decline) in their liquid assets—cash. Both perceptions are rates of change. These two perceptions are sometimes in conflict. To see your investments grow, you need to get into the market and buy. To increase your supply of cash, you need to sell some of your investments. You can’t have it both ways at once, but both goals are desirable. PCT Model Above the lighter dotted line: Our two control systems Between the two dotted lines: Perceptions controlled by lower-level systems (not explicitly modeled) Below the heavier dotted line: Variables and relationships in the external environment. We treat the market price of the investment as a disturbance. Here’s the price of the fictitious stock. It follows the ups and downs of the Dow-Jones average from 2000 to 2010. There is a 6% drop overall. 140.00 130.00 120.00 110.00 Dollars 100.00 90.00 80.00 70.00 60.00 50.00 40.00 1995 1996 1997 1998 1999 2000 2001 Date 2002 2003 2004 2005 2006 Here are the starting values for the robot investors. Each investor starts with $200,000 in assets: Stock worth $100,000 (1000 shares at $100 a share) and $100,000 in cash. Minimum transaction is 50 shares bought or sold. Each run of the simulation is 520 iterations (weeks). …see what they’re worth at the end of ten years. One example - Derek Derek’s profile: Control System Reference Value System Gain Investment Growth 15% 1000 Liquidity Protection 15% 1000 How did Derek do? 997 shares: $93,621.54 Cash: $110,953.38 Total assets: $204,574.92 2.3% GAIN 1,400 Market Price (times $10) 1,300 1,200 1,100 Shares Held 1,000 900 800 700 Cash on Hand (in $100's) 600 500 400 Dec 1995 Dec 1996 Dec 1997 Dec 1998 Dec 1999 Dec 2000 Dec 2001 Dec 2002 Dec 2003 Dec 2004 Dec 2005 Date How Derek did it He bought when the price was going down and sold when the price was going up. Contributions of this study to theories of economic behavior The agents demonstrate that an actor can appear “rational” without having any ability for rational deliberation. The findings call rational choice theory into question, since none of the assumptions for rational choice theory are satisfied by the robot investors. PCT offers an intriguing alternative to the received wisdom about economic behavior. Emergent Group Processes Computer simulations of multiple agents Each formed from control systems with a reference and gain for a variable – e.g. proximity to others McPhail, Powers & Tucker (1992) - demo Simulation ‘Run’ A 14 SA’s were given 2 identical goals:1) avoid collisions with anything in their path 2) pursue the target actor (PT) Co-ordinates of origin for SA’s OUTCOME: Ring Formation Co-ordinates of origin of target Simulation ‘Run’ B Randomly distributed obstacles introduced Co-ordinates of origin for SA’s Stationary Actors OUTCOME: Ring formation Effectively guided round obstacles Atypicality of Runs A & B The formations evident in Runs A & B were atypical for 2 reasons: 1) The rings formed were very symmetrical, and this is unlikely to happen in the non-simulated world. Almost perfect circle 2) Most gatherings are not made up of solitary individuals, rather they are made up of ‘companion clusters’ such as families or friendship groups. Therefore, the next run was programmed differently. All SA’s (except the target actor) were programmed to follow another actor, to make asymmetric pairs, with the latter member of each pair pursuing PT (the target). [ SA1 SA2 PT] Simulation ‘Run’ C & D Run C (with asymmetrical pairs) ended in a double arc form. Again this is rarely observed in public places. Run D included 15 obstacles (stationary actors) and this caused some of the pairs to get separated. This resulted in a less symmetrical, more authentic ring that better approximates social forms observed in the real world. Modelling Social Interactions (Mansell et al., in prep) McPhail et al. used quantitative data to generate qualitative outcomes Can a PCT model be used to generate quantitative models of actual behaviour? Personal Distance Paradigm The Method • 45 participants • 5 participants each trial • Each participant pairs up for a conversation with each of the other 4 participants • Video camera used to record personal distance for each pairing The conversation task analysis – pairings of participants as A or B Participant 1 2 3 4 5 1 2 A 3 4 5 B B A A B B A B A Analysis Computer model of two feedback loops controlling the same input (personal distance) Estimate Reference Value (R) from the mean of the distance in A pairings Use trial & error changes to select Rs and Gains (G) that generate the measured distance as their input Input values of R and G into model to simulate novel pairings Reference Value The PCT computer program (Mansell et al, 2011) Referenc Output e Gain Value -Has all the functions, signals and amplification factor of gain (Input and output gain) -Adjust the reference value -Adjust the gain -2 control systems -Until desired input achieved = measured personal distance Referenc Input e Quantity Value Hypothesis The estimated personal distance generated by computer models trained on A pairings will generate an estimate of personal distance for B pairings that is closely correlated with actual personal distance r2=0.32 The computer model (r = .563, n= 44, p <0.01) This is close to the correlation between actual personal distance between A and B pairings: r = 0.60 Modelling Psychopathology PCT states that chronic psychological distress is caused by unresolved conflict between control systems (Powers, 1973) Control systems are organised in a hierarchy to manage complex goals Psychological change is carried out through a trial-and-error learning process called reorganisation Awareness and therefore learning needs to be directed to the deeper, superordinate systems Explanation of these components… SYSTEM CONCEPT Be a responsible person PRINCIPLE Follow through on commitments PROGRAM Drive over and return the notes RELATIONSHIP “Drivingness” SEQUENCE Make a turn TRANSITION Turning of steering wheel CONFIGURATION Fingers around rim of wheel SENSATION Gripping INTENSITY Muscle tensions INPUT Effect on environment MULTI-LEVEL CONTROL SYSTEM Adapted from Carver & Scheier (1981, 1982) based on Powers (1973) Control Hierarchies: - Organisations of negative feedback loops that specify goals for lower level systems - required for complex control Hierarchies in PCT Robots Kennaway (1999) ◦ ◦ ◦ ◦ Archy two level hierarchy angles; orientation simulated 3D landscape Lippett (2005) – Real world version of Archy Young (2011) – controls perception of closeness; two levels Neal et al. (1997) – robotic grasping of fish fingers on production line Reorganisation as Learning Error is assessed within intrinsic systems, e.g. related to survival Discrepancy sensed between current error and ‘intrinsic error’ This discrepancy drives random variation in the change in properties of the control systems until error is reduced New change in organisation persists No specific behaviour is learned; not reinforcement Modelled on computer simulations (Marken & Powers, 1989; Powers, 2008) Chronic conflict requires reorganisation of the systems that set the standards for the conflicting systems – i.e. prioritisation in upper level systems (see also social construal theory) PCT Model of Conflict (Carey,2006) Modelled in computer simulation (Akgonul, Mansell, Li, Powers, & Carey) Compare error reduction with reorganisation in mid vs. high level systems PCT and Spoken Language Contemporary automated speech and machine generated speech poor Direct contrast to the robust and flexible nature of human speech. PRESENCE “PREdictive SENsorimtor Control and Emulation” (Moore, 2007) ◦ based on PCT principles ◦ novel approach for linguistics. ◦ Move away from independent study e.g. speech recognition, spoken language. Evidence to support PCT in spoken language ◦ Adjust volume according to surroundings. ◦ Problems when voice cannot be heard as expected. Demonstrating PRESENCE Moore (2007) built a Lego NXT computer model (Alpha Rex) to maximize synchrony of its own behaviour with an external source. ◦ To copy the rhythm of speech 3 control loops: 1. Generate own rhythmic behaviour 2. Sense its own sounds 3. Sense external sounds Tapping Rhythm Alpha Rex reduces the error Alpha Rex predicts next word Implications A simple model can produce adaptive behaviour that would be very difficult to program explicitly. ◦ 3 control loops only The model shows that behaviour can be learned and provide anticipatory ability ◦ the timing of next word The model shows that PCT can be used to make predictions about behaviour. ◦ Allows for use of context in spoken language ◦ Simulation & Learning (Moore, 2007) Text-Based Communication for Therapy Manage Your Life On-line MYLO (Manage Your Life On-line) ◦ aims to simulate Method of Levels as described by Carey (2006; 2008). Questioning technique to sustain awareness at higher levels in a control system hierarchy Developed at University of Manchester for improving access to self-help Evaluated in comparison to a control condition (ELIZA; Weisenbaum, 1969) Promising results on self-reported distress… Manage Your Life On-line You said: I'm worried about graduating next year. I want to get a job but my parents are putting pressure on me to do a masters. MYLO: What do you think about the pressure that's being applied? You said: It's stressing me out. I don't think that a masters is for me, but they keep going on and on. MYLO: How did you work that out? You said: I have been struggling a little since I started the third year and don't know if I am even going to be able to pass this! I'm worried about failing. MYLO: You mentioned a feeling - ‘worried' - how is that right now? Question You mentioned a feeling - ‘worried' - how is that right now? Answer: It's pretty bad. I have an essay to write and I don't really understand it. I'm so stupid. submit How MYLO works I have been feeling [[worried] about doing] a presentation at uni for my dissertation. [I know that] I’m going to forget what to say and look stupid. • • • • Anxious Worried Scared etc… • • • • I know that I think that I’m sure he etc… • • • • next time he might think Worried about doing Etc… Perception Sensing Future Events +55 more • Can you tell me more about what “worried” is like for you? • When you say “worried” how does that actually feel for you? • What do you think about feeling '*'? • What makes you think that? • How does thinking this affect you? • What makes you believe this is the case? • How is it to picture the future like that just now? • How does picturing the future like that make you feel now? • How does that image sit with you? Summary of Current Research Compare effects of MYLO with control condition (ELIZA) on distress ◦ Develop as a mental health intervention Computer models of personal space control ◦ Hierarchical goals in future Computer models of psychopathology and psychotherapy ◦ Build algorithm that automatically selects the most adaptive focus of reorganisation ◦ Model two agents to illustrate how control is critical in psychotherapy See www.pctweb.org Some Points for Discussion… How does PCT compare to existing models? Where to get software – most online at www.billpct.org/ ; also liaising with Powers, Marken, McClelland, Moore, etc… Need an empirical overview paper Does this truly integrate disciplines?