Action Selection for Hammer Shots in Curling
Authors: Zaheen Farraz Ahmad, Robert C. Holte, Michael Bowling
IJCAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments using our curling physics simulator show that the adapted Delaunay triangulation s shot selection outperforms other algorithms, and with some caveats, exceeds Olympic-level human performance. |
| Researcher Affiliation | Academia | Zaheen Farraz Ahmad, Robert C. Holte, Michael Bowling, Department of Computing Science, University of Alberta {zfahmad, rholte, mbowling}@ualberta.ca |
| Pseudocode | Yes | DS s first stage proceeds as follows: 1. Sample the (stochastic) objective function at points distributed uniformly over the range of values of the action parameters. 2. Apply Delaunay triangulation [Lee and Schachter, 1980] to the sampled points to partition the continuous action space into a set of disjoint regions. 3. Assign each region a weight, which we discuss below. 4. Sample the set of regions with replacement, with the probability of a region proportional to its weight. 5. Each time a region is selected in the previous step, uniformly randomly sample a point within it and add it to the set of sampled points, with its value sampled from the (stochastic) objective function at that point. 6. Repeat from step 2 for a fixed number (T) of iterations. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing its source code or a link to a code repository. |
| Open Datasets | Yes | The data used came from http://curlingzone.com, and included both women s and men s tournaments, although almost no difference was observed when restricting to data only from one gender or when including only championship level events. |
| Dataset Splits | No | A set of 397 hammer shot states from these logs were used in the parameter sweeps mentioned below. A separate set of 515 hammer shot states (the test states) from these logs were used to evaluate the systems (including the humans). |
| Hardware Specification | No | Computing resources were provided by Compute Canada and Calcul Qu ebec. However, no specific hardware details (e.g., GPU/CPU models, memory) are provided. |
| Software Dependencies | No | The curling simulator used in this paper is implemented using the Chipmunk 2D rigid body physics library with an artificial lateral force that visually recreates empirically observed stone trajectories and modified collision resolution to visually match empirically observed elasticity and energy conservation when rocks collide. |
| Experiment Setup | Yes | The parameter sweep for Delaunay Sampling (DS) chose a value of 14 for σ. This budget includes 100 samples for initializing the triangulation over each turn, 100 samples per iteration of the first stage, and 100 samples for the final UCB stage. The parameters for HOO described by Bubeck et al. [2009] were set by a parameter sweep to = 1 p 2 and UCB constant C = 0.01. PSO s parameters were set to c1 = 1.5, c2 = 1.5, w = 0.7 and 50 particles. We set the initial standard deviation to σ0 = 0.25 and normalized the action parameters and v to range between 0 and 1. The parameters for step-size control and the covariance matrix adaptation were set in accordance to the values recommended by Hansen [2016]. We chose the squared-exponential function as our GP kernel and set the noise variance, σn = 0.01. |