Recent advances in applying deep learning to planning have shown that Deep Reactive Policies (DRPs) can be powerful for fast decision-making in complex environments. However, an important limitation of current DRP-based approaches is either the need of optimal planners to be used as ground truth in a supervised learning setting or the sample complexity of high-variance policy gradient estimators, which are particularly troublesome in continuous state-action domains. In order to overcome those limitations, we introduce a framework for training DRPs in continuous stochastic spaces via gradient-based policy search. The general approach is to explicitly encode a parametric policy as a deep neural network, and to formulate the probabilistic planning problem as an optimization task in a stochastic computation graph by exploiting the re-parameterization of the transition probability densities; the optimization is then solved by leveraging gradient descent algorithms that are able to handle non-convex objective functions. We benchmark our approach against stochastic planning domains exhibiting arbitrary differentiable nonlinear transition and cost functions (e.g., Reservoir Control, HVAC and Navigation). Results show that DRPs with more than 125,000 continuous action parameters can be optimized by our approach for problems with 30 state fluents and 30 action fluents on inexpensive hardware under 6 minutes. Also, we observed a speedup of 5 orders of magnitude in the average inference time per decision step of DRPs when compared to other state-of-the-art online gradient-based planners when the same level of solution quality is required.

When modeling real-world decision-theoretic planning problems in the Markov decision process (MDP) framework, it is often impossible to obtain a completely accurate estimate of transition probabilities. For example, natural uncertainty arises in the transition specification due to elicitation of MDP transition models from an expert or data, or non-stationary transition distributions arising from insufficient state knowledge. In the interest of obtaining the most robust policy under transition uncertainty, the Markov Decision Process with Imprecise Transition Probabilities (MDP-IPs) has been introduced to model such scenarios. Unfortunately, while solutions to the MDP-IP are well-known, they require nonlinear optimization and are extremely time-consuming in practice. To address this deficiency, we propose efficient dynamic programming methods to exploit the structure of factored MDP-IPs. Noting that the key computational bottleneck in the solution of MDP-IPs is the need to repeatedly solve nonlinear constrained optimization problems, we show how to target approximation techniques to drastically reduce the computational overhead of the nonlinear solver while producing bounded, approximately optimal solutions. Our results show up to two orders of magnitude speedup in comparison to traditional "flat" dynamic programming approaches and up to an order of magnitude speedup over the extension of factored MDP approximate value iteration techniques to MDP-IPs.

Planning to reach a goal is an essential capability for rational agents. In general, a goal specifies a condition to be achieved at the end of the plan execution. In this article, we introduce nondeterministic planning for extended reachability goals (i.e., goals that also specify a condition to be preserved during the plan execution). We show that, when this kind of goal is considered, the temporal logic ctl turns out to be inadequate to formalize plan synthesis and plan validation algorithms. This is mainly due to the fact that the ctl's semantics cannot discern among the various actions that produce state transitions. To overcome this limitation, we propose a new temporal logic called α-ctl. Then, based on this new logic, we implement a planner capable of synthesizing reliable plans for extended reachability goals, as a side effect of model checking.

No abstract