Remarkably, we can do so even when those actions lead to failure, enabling us to assist others when we detect that they might not achieve their goals. How might we endow machines with similar capabilities? Here we present an architecture capable of inferring an agent's goals online from both optimal and non-optimal sequences of actions. These models are specified as probabilistic programs, allowing us to represent and perform efficient Bayesian inference over an agent's goals and internal planning processes. To perform such inference, we develop Sequential Inverse Plan Search (SIPS), a sequential Monte Carlo algorithm that exploits the online replanning assumption of these models, limiting computation by incrementally extending inferred plans as new actions are observed.

Weaknesses: My main concerns for the work are about specific assumptions made regarding the agent's planning algorithm and how close the effectiveness of the goal recognition system is tied to having access to the specific planning algorithm and parameters used by the agent generating the observations. I would have liked to see experimental results that at least shows some level of robustness of the system towards mismatch between the planning algorithm used by the goal recognition system and the method used to generate the observations for the study. Below I have provided a more detailed discussion of my main concerns Specific Algorithm Used: The paper makes some specific assumptions on the kind of algorithm that could be used to simulate the bounded decision making. I see no reason to believe that this is general enough to capture behavior of any arbitrary resource bounded decision-maker (for example consider one that is quite similar to the one discussed, but is also memory bounded and can only hold limited possible nodes in its open list) or that this is in anyways similar to how a human would make such decisions (which is important if the primary goal is to be able to predict human goals). While the paper notes that people use heuristics as well, those may be quite different from the ones that are popular in planning literature.

Remarkably, we can do so even when those actions lead to failure, enabling us to assist others when we detect that they might not achieve their goals. How might we endow machines with similar capabilities? Here we present an architecture capable of inferring an agent's goals online from both optimal and non-optimal sequences of actions. These models are specified as probabilistic programs, allowing us to represent and perform efficient Bayesian inference over an agent's goals and internal planning processes. To perform such inference, we develop Sequential Inverse Plan Search (SIPS), a sequential Monte Carlo algorithm that exploits the online replanning assumption of these models, limiting computation by incrementally extending inferred plans as new actions are observed.