Ricardo Aler

EAAI Journal 2022 Journal Article

Direct estimation of prediction intervals for solar and wind regional energy forecasting with deep neural networks

• Antonio Alcántara
• Inés M. Galván
• Ricardo Aler

Deep neural networks (DNN) are becoming increasingly relevant for probabilistic forecasting because of their ability to estimate prediction intervals (PIs). Two different ways for estimating PIs with neural networks stand out: quantile estimation for posterior PI construction and direct PI estimation. The former first estimates quantiles, which are then used to construct PIs, while the latter directly obtains the lower and upper PI bounds by optimizing some loss functions, with the advantage that PI width is directly considered in the optimization process and thus may result in narrower intervals. In this work, two different DNN-based models are studied for direct PI estimation, and compared with DNN for quantile estimation in the context of solar and wind regional energy forecasting. The first approach is based on the recent quality-driven loss and is formulated to estimate multiple PIs with a single model. The second is a novel approach that employs hypernetworks (HN), where direct PI estimation is formulated as a multi-objective problem, returning a Pareto front of solutions that contains all possible coverage-width optimal trade-offs. This formulation allows HN to obtain optimal PIs for all possible coverages without increasing the number of network outputs or adjusting additional hyperparameters, as opposed to the first direct model. Results show that prediction intervals from direct estimation are narrower (up to 20%) than those of quantile estimation, for target coverages 70%–80% for all regions, and also 85%, 90%, and 95% depending on the region, while HN always achieves the required coverage for the higher target coverages.

IJCAI Conference 2007 Conference Paper

• Susana Fern
• aacute; ndez
• Ricardo Aler
• Daniel Borrajo

As any other problem solving task that employs search, AI Planning needs heuristics to efficiently guide the problem-space exploration. Machine learning (ML) provides several techniques for automatically acquiring those heuristics. Usually, a planner solves a problem, and a ML technique generates knowledge from the search episode in terms of complete plans (macro-operators or cases), or heuristics (also named control knowledge in planning). In this paper, we present a novel way of generating planning heuristics: we learn heuristics in one planner and transfer them to another planner. This approach is based on the fact that different planners employ different search bias. We want to extract knowledge from the search performed by one planner and use the learned knowledge on another planner that uses a different search bias. The goal is to improve the efficiency of the second planner by capturing regularities of the domain that it would not capture by itself due to its bias. We employ a deductive learning method (\ebl) that is able to automatically acquire control knowledge by generating bounded explanations of the problem-solving episodes in a Graphplan-based planner. Then, we transform the learned knowledge so that it can be used by a bidirectional planner.

ICAPS Conference 2002 Conference Paper

On Control Knowledge Acquisition by Exploiting Human-Computer Interaction

• Ricardo Aler
• Daniel Borrajo

In the last decade, there has been a strong and increasing interest on building fast planning systems. From it, we have seen an enormous improvement in relation to the solvability horizon of current planning techniques. Most of these techniques rely on the translation of the predicate logic description of domains into propositional representations of them. Then, a fast search procedure is applied for obtaining solutions to planning problems. While this is an important step towards solving the planning problem, we believe older planners that are based on predicate logic computation can still be competitive when they are combined with learning capabilities. This is so, because they can acquire and use more abstract representations of control knowledge (CK) which are easier to describe and maintain by humans and/or automatic systems than the ones based on propositional logic. In this paper, we present the results we have obtained with a relatively "old" planner, Prodigy4. 0, powered with CK that has been acquired using a mixed human-computer collaboration. We advocate for the initial generation of CK by a learning system, and its later refinement by a human. In fact, we can iterate this process until the desired result has been achieved. We show results in the logistics domain used at AIPS'00 that are at the same level when compared with other techniques in the Track on hand-tailored planning systems (Track2).

AIJ Journal 2002 Journal Article

Using genetic programming to learn and improve control knowledge

• Ricardo Aler
• Daniel Borrajo
• Pedro Isasi

The purpose of this article is to present a multi-strategy approach to learn heuristics for planning. This multi-strategy system, called Hamlet-EvoCK, combines a learning algorithm specialized in planning (Hamlet) and a genetic programming (GP) based system (EvoCK: Evolution of Control Knowledge). Both systems are able to learn heuristics for planning on their own, but both of them have weaknesses. Based on previous experience and some experiments performed in this article, it is hypothesized that Hamlet handicaps are due to its example-driven operators and not having a way to evaluate the usefulness of its control knowledge. It is also hypothesized that even if Hamlet control knowledge is sometimes incorrect, it might be easily correctable. For this purpose, a GP-based stage is added, because of its complementary biases: GP genetic operators are not example-driven and it can use a fitness function to evaluate control knowledge. Hamlet and EvoCK are combined by seeding EvoCK initial population with Hamlet control knowledge. It is also useful for EvoCK to start from a knowledge-rich population instead of a random one. By adding the GP stage to Hamlet, the number of solved problems increases from 58% to 85% in the blocks world and from 50% to 87% in the logistics domain (0% to 38% and 0% to 42% for the hardest instances of problems considered).

ICML Conference 2000 Conference Paper

Knowledge Representation Issues in Control Knowledge Learning

• Ricardo Aler
• Daniel Borrajo
• Pedro Isasi

ICML Conference 1998 Conference Paper

Genetic Programming and Deductive-Inductive Learning: A Multi-Strategy Approach

• Ricardo Aler
• Daniel Borrajo
• Pedro Isasi

ICAPS Conference 1998 Conference Paper

Genetic Programming of Control Knowledge for Planning

• Ricardo Aler
• Daniel Borrajo
• Pedro Isasi

It is well knownthat planning for complex domainsand non-trivial problemsis a difficult task for domain independent planners. Adding domain knowledgeto the planner is a simple approach to improveits efficiency for a given domain. In this paper we present EvoCK, a system based on Genetic Programming, which produces control knowledge for a nonlinear domain independent planner (PRODIGY4. 0). EvoCKhas been tested in two domains (the blocksworld and logistics), and very good experimental results have been obtained both in terms of the numberof problems solved and of the quality of the control-rules produced.