"Computational rationality as a theory of interaction" - CHI'22 paper
Download as PPTX, PDF0 likes128 views
This document discusses computational rationality as a theory of interaction. It proposes that interaction emerges from an agent acting within internal and external bounds based on its bounded cognition and environment. Reinforcement learning agents maximize rewards through interaction, but computational rationality models combine reinforcement learning with cognitive architectures. The agent acts via its internal environment and is boundedly optimal. The theory provides commitments for modeling human-computer interaction tasks, and has been applied to behaviors like typing, multitasking, and decision making. The goal is to accurately predict human behavior and explain adaptation. Significant progress has been made in applying these models in HCI, but opportunities remain to model motivations, learning, social aspects, and design interventions.
![Computational Rationality as a Theory of Interaction
Opportunity: Explain behavior
[Kangasrääsiö et al. CHI’17, Cognitive Science 2019]
Inverse modeling with likelihood-free inference Bayesian methods
Examples in the paper](https://image.slidesharecdn.com/chi22computationalrationality-220422135941/85/Computational-rationality-as-a-theory-of-interaction-CHI-22-paper-15-320.jpg)
"Computational rationality as a theory of interaction" - CHI'22 paper
- 1. Computational Rationality as a Theory of Interaction Antti Oulasvirta, Jussi Jokinen, Andrew Howes
- 2. Computational Rationality as a Theory of Interaction Interactive behavior is highly adaptive
- 3. Computational Rationality as a Theory of Interaction Interaction emerges Actions Thoughts Beliefs Goals Capabilities Design
- 4. Computational Rationality as a Theory of Interaction Credit: Alex Jung / Aalto University RL agents adapt via interaction But AI research has no interest in explaining human adaptation Agents try to learn sequential action policies that maximize rewards
- 5. Computational Rationality as a Theory of Interaction Example: Adaptation in typing Jokinen et al. CHI’21 Write fast w/out error Adapts when error occurs Turn on autocorrection No human data needed for training
- 6. Computational Rationality as a Theory of Interaction Increasing interest in HCI Text entry Multitasking Decision-making Theoretical commitments? Core modeling ideas?
- 7. Computational Rationality as a Theory of Interaction Combines RL with cognitive architectures Lewis et a al. 2014; Gershman et al. 2015; Lieder & Griffiths 2019 Computational Rationality Cognitive architectures Agents act via their bounded cognition Reinforcement learning Bounded optimality Interaction emerges as adaptation within internal and external bounds
- 8. Computational Rationality as a Theory of Interaction Cognitive Architectures Computational Rationality An information processor An agent Manually specified rule set Policy obtained via optimization Bounded by cognition Bounded by cognition Program POMDP Comparison with cognitive architecture models
- 9. Computational Rationality as a Theory of Interaction Theoretical commitments in HCI 1. Agent faces a stochastic sequential decision problem 2. Agent acts in the world via its internal environment 3. Bounds exist on information processing in the internal environment 4. Agent is boundedly optimal
- 10. Computational Rationality as a Theory of Interaction Agent structure point-of-view Standard RL Computational rationality
- 11. Computational Rationality as a Theory of Interaction Policy estimation view Lewis & Howes 2019 Subjective utility payoff surface Boundedly optimal policy Standard RL Computational rationality Behaviors admitted by the internal environment Behaviors admitted by the external environment Behaviors admitted by the external environment
- 12. Computational Rationality as a Theory of Interaction Modeling in practice External environment Agent Eye mov commands Finger mov commands Bounded vision Reward function Max speed and min errors Finger and eye positions Motor system Perception Internal environment Eye mov Finger mov Generates moment-by- moment traces, from which summary statistics can be computed
- 13. Computational Rationality as a Theory of Interaction Comparison with human data Jokinen et al. CHI’21 Human data Model predictions The goal in CR is to accurately predict human behavior
- 14. Computational Rationality as a Theory of Interaction Significant progress last 5 years 147 papers found 87 peer-reviewed 20 models of HCI tasks 15 used a rigorous method to find an optimal policy Six interactive tasks • Pointing • Menu selection • Driving • Typing • Multitasking • Decision-making
- 15. Computational Rationality as a Theory of Interaction Opportunity: Explain behavior [Kangasrääsiö et al. CHI’17, Cognitive Science 2019] Inverse modeling with likelihood-free inference Bayesian methods Examples in the paper
- 16. Computational Rationality as a Theory of Interaction Opportunity: Design interventions Examples in the paper
- 17. Computational Rationality as a Theory of Interaction AI HCI’s interest is distinct HCI Supporting user- centered design and engineering Cognitive science Human intelligence General intelligence
- 18. Computational Rationality as a Theory of Interaction Agenda: From micro- to macro-HCI Motivational dynamics Human learning Situations Social interaction
- 19. Computational Rationality as a Theory of Interaction Antti Oulasvirta, Jussi Jokinen, Andrew Howes