Computational Causal Behaviour Models (CCBM)

Computational Causal Behaviour Models (CCBM) are a modelling approach describing human behaviour in terms of precondition-effect rules, different action selection heuristics and probabilistic action durations. The CCBM toolkit allows to compile these models into various Bayesian filters such as Hidden Markov Models, Particle Filter or Marginal Filter. This allows for CCBM to reason about the user’s situation, actions, intentions, and causes behind the observed behaviour based on noisy observations.

Details

Computational state space models (CSSMs) enable the knowledge-based construction of Bayesian filters for recognizing intentions and reconstructing activities of human protagonists in application domains such as smart environments, assisted living, or security. Computational, i. e., algorithmic, representations allow the construction of increasingly complex human behaviour models. Computational Causal Behaviour Models (CCBM) are a modelling approach that represent human behaviour computationally by means of preconditions and effects, different action selection heuristics and probabilistic action durations. The CCBM toolkit compiles these models into various Bayesian filters such as Hidden Markov Models, Particle Filter or Marginal Filter. The Marginal Filter is specifically tailored for categorical state spaces, as they are generated by CCBM. It has been shown to outperform the standard method for approximative Bayesian inference – the Particle Filter. This allows for CCBM to reason about the user’s situation, actions, intentions, and causes behind the observed behaviour based on noisy and ambiguous observations such as sensor data.

Graphical model used in CCBM
The graphical model used in CCBM. Nodes with single outline denote hidden random variables, whereas nodes with double outline denote observable random variables. Boxes represent tuples of variables. Edges represent dependencies of a (tuple of) random variable(s) (edge ending) from another (tuple of) random variable(s) (edge starting).

The dynamic Bayesian network used as probabilistic model in the CCBM approach is defined as follows: The hidden state Xt is a five-tuple (At, Dt, Gt, St, Ut). The action executed at time t is denoted by At, Dt is a boolean flag signalling whether At-1 is terminated in the left-open and right-closed interval (t-1, t] and a new action At has to be selected. Ut is the starting time of the action At. The variable St is the current environment state and Gt denotes the goal (or intention) of the user at time t. The observation Yt=(Wt, Zt) for time-step t consists of two conditional independent parts, the state observation Wt and the action observation Zt. Vt is the timestamp of the observation data sequence.

Key Publications