Human Goal Recognition as Bayesian Inference: Investigating the Impact of Actions, Timing, and Goal Solvability

TL;DR

The paper reframes goal recognition as Bayesian inference, incorporating actions, timing, and goal solvability within a Sokoban domain to understand human inferences. It decomposes the likelihood into timing and action components and uses a solvability-aware Adaptive Lookahead Planner to estimate likelihoods, with solvability captured in the prior via an Easiness model. Empirical results show actions are typically the strongest cue, but timing and solvability influence inferences when actions are uninformative, and a Bayesian model with an Easiness prior and online likelihood best matches human judgments. The work advances human-like goal recognition, offering a principled, interpretable approach with potential applications in explainable AI and transparent planning. The Bayesian formulation is $P(G|O) \propto Prior(G)\cdot LL(O,G)$ with $LL(O,G)=LL_T(O,G)\cdot LL_A(O,G)$, and the solvability-aware planner provides a concrete mechanism for likelihood estimation. The Easiness prior captures domain-independent priors over goal solvability and difficulty, and the online likelihoods based on planner simulations align closely with observed human behavior across map types. Overall, the study demonstrates that a human-centered Bayesian model can outperform traditional action-only recognizers and invites further refinement of planning-based likelihoods for real-time, interpretable goal inference.

Abstract

Goal recognition is a fundamental cognitive process that enables individuals to infer intentions based on available cues. Current goal recognition algorithms often take only observed actions as input, but here we use a Bayesian framework to explore the role of actions, timing, and goal solvability in goal recognition. We analyze human responses to goal-recognition problems in the Sokoban domain, and find that actions are assigned most importance, but that timing and solvability also influence goal recognition in some cases, especially when actions are uninformative. We leverage these findings to develop a goal recognition model that matches human inferences more closely than do existing algorithms. Our work provides new insight into human goal recognition and takes a step towards more human-like AI models.

Figures (4)

• Figure 1: Examples showing three types of Sokoban maps in which action, timing and solvability might affect human goal inference. After performing two forced moves indicated by the purple arrows, the actor executes a key step indicated by the pink arrow. Complete sets of maps used in our experiment can be found in the supplementary materials. (a) An action map. The red goal is achievable and the green goal is not, and the actor moves left at the key step. (b) An easy-goal map. The red goal is easy to achieve but the green goal is not achievable. The key move (not shown) involves a push to the left. (c) A second easy-goal map. The red goal is easy to achieve but the path to the green goal is more complex. At the key move (again not shown) the actor pushes the box to the left. (d) A competing-path map. There is one good path (red arrows) to the red goal and two good paths (green arrows) to the green goal. The actor moves up at the key step.
• Figure 2: Results for the planning phase. (a) Proportion of participant choices for the action in action maps. Cons means consistent with our manipulation in the goal recognition phase. The model employs softmax action selection with a temperature parameter set to 5. (b) Average Planning time for easy and hard goals in easy-goal maps. The effect of thinking time is significant for both humans and the model ($p<0.001$). For both (b) and (c), error bars show the standard deviation of the mean planning time (measured in seconds). (c) Average Planning time for competing and no-competing goals in competing-path maps. The effect of thinking time is significant for both human and model ($p<0.05$). (d) Number of steps taken in unsolvable instances for humans (x-axis) and the model (y-axis). Human responses and model predictions are strongly correlated ($r(7)=0.65, p=0.05$).
• Figure 3: Results for prior instances in the goal-recognition phase. (a) Response distribution for prior instances where goal A is solvable and goal B is not. Blue bars indicate a preference for solvable goal A while red bars represent a preference for unsolvable goal B. (b) Comparison between human responses and the easiness model. The x-axis represents the model's predicted probability of choosing the easy goal, and the y-axis represents the human prior observed in the experiment. The instances are represented as circles, crosses or stars based on whether neither, one or both goals are unsolvable. (c) Response distribution from panel (a) broken down by the three subtypes.
• Figure 4: Comparison between model predictions and human inferences. All model labels show the prior followed by the likelihood: for example, uniform + emp is the model with uniform Prior and the empirical likelihood. emp(a) and online(a) are likelihoods that incorporate actions but not timing information. For readability, log likelihoods (higher is better) are shown as offsets relative to the log likelihood of the uniform+offline model.

Theorems & Definitions (1)

• definition 1