Folk Psychology forHuman Modelling :Folk Psychology forHuman Modelling Extending the BDI Paradigm Emma NorlingDepartment of Computer Science and Software EngineeringThe University of Melbourne


Outline :Outline Human Modelling, Folk Psychology and BDI Extending BDI for Human Modelling Perception and action Decision making Conclusions & future work


Human Modelling :Human Modelling Human modelling is the simulation of human behaviour Includes applications such as Operations analysis Human-in-the-loop training Entertainment Focus is on “holistic” behaviour


Folk Psychology :Folk Psychology Folk psychology is ‘layman’s psychology’— a means of explaining the behaviour of others via their reasoning It does not attempt to capture how the mind really works It is a robust mechanism for explaining and predicting the behaviours of others It is also commonly used to explain one’s own behaviour


The BDI Agent Architecture :AGENT The BDI Agent Architecture Sensors Actions Based in folk psychology – Dennett’s intentional stance The agent will intend to do what it believes will achieve its goals


BDI and Human Modelling :BDI and Human Modelling BDI has been used with considerable success for human modelling The success is largely due to its folk psychological roots Facilitates knowledge representation Agents’ behaviour can be understood by laymen However…


BDI and Human Modelling :BDI and Human Modelling The intentional stance is a high level abstraction of human reasoning It ignores many generic aspects of human behaviour, such as emotion, fatigue, mental and physical limitations More of these should be included in a framework designed for human modelling


A BDI-BasedHuman Modelling Framework :A BDI-BasedHuman Modelling Framework Given that BDI has had success in human modelling, can it be extended to be better-suited to this purpose? Proposed approach: integrate folk psychological explanations where applicable, attach modules otherwise


An attached module :An attached module Perception and action


Attaching a Module :Attaching a Module Vision and motion operates at such a low level that folk psychology is not typically used to describe them Vision module feeds inputs to the agent’s beliefs; action module introduces timing and errors to the agent’s action


An Integrated Explanation :An Integrated Explanation Recognition Primed Decision-making (RPD)


An Integrated Explanation :An Integrated Explanation BDI uses utility-based decision-making, or selects first applicable, but humans use a wide range of decision strategies In certain types of domain, RPD can account for up to 95% of the choices RPD is a descriptive model of decision making, using folk psychological concepts


Implementation :Implementation Agent must be able to learn to recognise situations in order to make decisions BDI agents use plan context to limit the plans considered, but this is static Use meta-level reasoning to enable Q-learning based plan selection Update Q-values on plan completion


Exploration of Parameters :Exploration of Parameters Q(st, at) ← Q(st, at) + α[rt+1 + γ max Q(st+1, a) Q(st, at)]


A Bigger Problem: Quake 2 :A Bigger Problem: Quake 2 BDI-based agents to play Quake 2 Connect to the game server as if they were a player client Receive the same data as the player client, send commands via simulated key presses and mouse actions Model expert Quake 2 players It doesn’t work! (But not because of the extension)


Conclusions :Conclusions To extend the BDI framework to better suit human modelling: Include more of the generic aspects of human behaviour/reasoning in the framework Maintain the folk psychological roots, because this is its greatest strength Integrate folk psychological explanations where they exist Use low-level explanations that do not impact on the knowledge representation otherwise


Future Work :Future Work COJACK – an extension to JACK enabling cognitively plausible models of human variability Social aspects of agency Can folk psychology be used to capture social intelligence?